Daily Papers

Scene understanding is central to general physical intelligence, and video is a primary modality for capturing both state and temporal dynamics of a scene. Yet understanding physical processes remains difficult, as models must combine object localization, hand-object interactions, relational parsing, temporal reasoning, and step-level procedural inference. Existing benchmarks usually evaluate these capabilities separately, limiting diagnosis of why models fail on procedural tasks. We introduce BARISTA, a densely annotated egocentric dataset and benchmark of 185 real-world coffee-preparation videos covering fully automatic, portafilter-based, and capsule-based workflows. BARISTA provides verified per-frame scene graphs linking persistent object identities to masks, tracks, boxes, attributes, typed relations, hand-object interactions, activities, and process steps. From these graphs, we derive zero-shot language-based tasks spanning phrase grounding, hand-object interaction recognition, referring, activity recognition, relation extraction, and temporal visual question answering. Experiments reveal strong variation across task families and no consistently dominant model family, positioning BARISTA as a challenging diagnostic benchmark for procedural video understanding. Code and dataset available at https://huggingface.co/datasets/ramblr/BARISTA.

Tool-using agents are increasingly expected to operate across realistic professional workflows, where they must interpret multimodal inputs, coordinate external tools, inspect intermediate artifacts, and revise their actions before producing a final result. Existing benchmarks, however, often evaluate tool use, computer use, and multimodal reasoning in isolation, leaving a gap between benchmark settings and end-to-end omni-modal tool use in the real world. To address this gap, we introduce MM-ToolBench, a benchmark and evaluation harness for task-oriented omni-modal tool use. MM-ToolBench contains 100 executable tasks from two macro task families, Customer Service and Intelligent Creation, covering 20 subcategory slices and supported by 27 MCP servers with 324 tools. The central design of MM-ToolBench is closed-loop multimodal verification: agents must execute tools, inspect rendered or transformed artifacts, and self-correct when outputs fail task-specific requirements. To make such evaluation scalable and verifiable, MM-ToolBench couples MCP-based execution with task-specific grounded evaluators and a semi-automated construction pipeline for scenario discovery, task instantiation, evaluator synthesis, and human audit. Experiments on 15 contemporary agentic models show that MM-ToolBench remains highly challenging: Claude Opus 4.6, commonly regarded as one of the strongest coding-agent models, achieves only 32.0% task success, far below the 94.0% human benchmark. We envision MM-ToolBench as a practical foundation for evaluating and advancing next-generation omni-modal tool-using agents through closed-loop multimodal verification.

Despite the recent success of multi-task learning and transfer learning for natural language processing (NLP), few works have systematically studied the effect of scaling up the number of tasks during pre-training. Towards this goal, this paper introduces ExMix (Extreme Mixture): a massive collection of 107 supervised NLP tasks across diverse domains and task-families. Using ExMix, we study the effect of multi-task pre-training at the largest scale to date, and analyze co-training transfer amongst common families of tasks. Through this analysis, we show that manually curating an ideal set of tasks for multi-task pre-training is not straightforward, and that multi-task scaling can vastly improve models on its own. Finally, we propose ExT5: a model pre-trained using a multi-task objective of self-supervised span denoising and supervised ExMix. Via extensive experiments, we show that ExT5 outperforms strong T5 baselines on SuperGLUE, GEM, Rainbow, Closed-Book QA tasks, and several tasks outside of ExMix. ExT5 also significantly improves sample efficiency while pre-training.

Multimodal Large Language Models have advanced visual reasoning, yet a purely textual chain of thought remains a bottleneck for questions that require fine-grained focus or view transformations. The ''think with images'' paradigm narrows this gap, but existing approaches are either constrained by fixed predefined toolkits or produce noisy intermediate images from unified multimodal methods. We pursue a third option: using a dedicated image editing model and decouple it with an understanding model. However, off-the-shelf image editors fail as reasoning assistants with two complementary gaps: a language-side gap, where editors trained as passive instruction-followers cannot map an abstract question to an appropriate visual transformation, and a generation-side gap, where edit correctness degrades as reasoning depth grows. Guided by this analysis, we introduce ETCHR (Editing To Clarify and Harness Reasoning), a question-conditioned, reasoning-aware image editor decoupled from the downstream understanding model and trained with a two-stage recipe targeted at the two gaps: Reasoning Imitation via supervised fine-tuning on edit trajectories, followed by Reasoning Enhancement with VLM-derived rewards for edit correctness and downstream reasoning accuracy. Since the editor is decoupled, ETCHR plugs into different open- and closed-source MLLMs in a training-free manner. Across five task families (fine-grained perception, chart understanding, logic reasoning, jigsaw restoration, and 3D understanding), ETCHR raises average Pass@1 from 55.95 to 60.77 (+4.82) with Qwen3-VL-8B, from 65.08 to 70.55 (+5.47) with Gemini-3.1-Flash-Lite, and from 76.55 to 81.16 (+4.61) with the 1T-parameter MoE model Kimi K2.5.

Vibe coding produces correct, executable code at speed, but leaves no record of the structural commitments, dependencies, or evidence behind it. Reviewers cannot determine what invariants were assumed, what changed, or why a regression occurred. This is not a generation failure but a control failure: the dominant artifact of AI-assisted development (code plus chat history) performs dimension collapse, flattening complex system topology into low-dimensional text and making systems opaque and fragile under change. We propose Agentic Consensus: a paradigm in which the consensus layer C, an operable world model represented as a typed property graph, replaces code as the primary artifact of engineering. Executable artifacts are derived from C and kept in correspondence via synchronization operators Phi (realize) and Psi (rehydrate). Evidence links directly to structural claims in C, making every commitment auditable and under-specification explicit as measurable consensus entropy rather than a silent guess. Evaluation must move beyond code correctness toward alignment fidelity, consensus entropy, and intervention distance. We propose benchmark task families designed to measure whether consensus-based workflows reduce human intervention compared to chat-driven baselines.

Multimodal large language models (MLLMs) demonstrate strong perception and reasoning performance on existing remote sensing (RS) benchmarks. However, most prior benchmarks rely on low-resolution imagery, and some high-resolution benchmarks suffer from flawed reasoning-task designs. We show that text-only LLMs can perform competitively with multimodal vision-language models on RS reasoning tasks without access to images, revealing a critical mismatch between current benchmarks and the intended evaluation of visual understanding. To enable faithful assessment, we introduce RSHR-Bench, a super-high-resolution benchmark for RS visual understanding and reasoning. RSHR-Bench contains 5,329 full-scene images with a long side of at least 4,000 pixels, with up to about 3 x 10^8 pixels per image, sourced from widely used RS corpora and UAV collections. We design four task families: multiple-choice VQA, open-ended VQA, image captioning, and single-image evaluation. These tasks cover nine perception categories and four reasoning types, supporting multi-turn and multi-image dialog. To reduce reliance on language priors, we apply adversarial filtering with strong LLMs followed by rigorous human verification. Overall, we construct 3,864 VQA tasks, 3,913 image captioning tasks, and 500 fully human-written or verified single-image evaluation VQA pairs. Evaluations across open-source, closed-source, and RS-specific VLMs reveal persistent performance gaps in super-high-resolution scenarios. Code: https://github.com/Yunkaidang/RSHR

While full-duplex speech agents enable natural, low-latency interaction by speaking and listening simultaneously, their consistency and task performance in multi-turn settings remain underexplored. We introduce Full-Duplex-Bench-v2 (FDB-v2), a streaming framework that integrates with an automated examiner that enforces staged goals under two pacing setups (Fast vs. Slow). FDB-v2 covers four task families: daily, correction, entity tracking, and safety. We report turn-taking fluency, multi-turn instruction following, and task-specific competence. The framework is extensible, supporting both commercial APIs and open source models. When we test full-duplex systems with FDB-v2, they often get confused when people talk at the same time, struggle to handle corrections smoothly, and sometimes lose track of who or what is being talked about. Through an open-sourced, standardized streaming protocol and a task set, FDB-v2 makes it easy to extend to new task families, allowing the community to tailor and accelerate evaluation of multi-turn full-duplex systems.

We introduce XTREME-S, a new benchmark to evaluate universal cross-lingual speech representations in many languages. XTREME-S covers four task families: speech recognition, classification, speech-to-text translation and retrieval. Covering 102 languages from 10+ language families, 3 different domains and 4 task families, XTREME-S aims to simplify multilingual speech representation evaluation, as well as catalyze research in "universal" speech representation learning. This paper describes the new benchmark and establishes the first speech-only and speech-text baselines using XLS-R and mSLAM on all downstream tasks. We motivate the design choices and detail how to use the benchmark. Datasets and fine-tuning scripts are made easily accessible at https://hf.co/datasets/google/xtreme_s.

As AI coding agents become more autonomous, they increasingly ship code iteratively, with the codebase persisting across sessions. This persistence creates a new attack surface: a misaligned or prompt-injected agent can distribute attacks across pull requests (PRs) and time its payload for the PR with the best natural cover. To study the resulting dynamics, we introduce Iterative VibeCoding, a setting for AI control, the study of safely deploying capable but potentially untrusted AI. In Iterative VibeCoding, a coding agent builds software over a sequence of PRs in a persistent codebase while pursuing a covert side task. Our benchmark includes two task families: CLI tools and Flask web services, across 20 total task variations. We use Claude Sonnet 4.5 as the attack agent and GPT-4o as the monitor. We compare gradual attacks, which distribute the side task across PRs, against non-gradual attacks concentrated in a single PR. No single monitor is robust to both: which strategy evades best (success while evading the monitor) depends on the monitor type, so a defender cannot close off both gradual and non-gradual attacks with any one monitor. High evasion (>= 65%) generalizes across model attack agent backends (Sonnet 4.5, Gemini 3.1 Pro, Kimi K2.5), confirming this is a property of the persistent-state attack surface rather than a single model's capability. Evasion also remains high across state-of-the-art monitor models and the gap between gradual and non-gradual evasion widens for more capable models. We introduce a stateful link-tracker monitor that tracks suspicious buildup across PRs. On both task families, it detects gradual attacks substantially better than diff monitors that merely see more accumulated history. Combining this stronger monitor with trajectory monitors in a four-monitor ensemble reduces gradual-attack evasion from 93% under the weakest standard diff monitor to 47%.

Executable evaluation -- checking the consequences of an agent's actions with a program rather than grading its prose -- has become a prominent way to assess tool-using AI agents in software settings. Electric power engineering has not yet had an analogous benchmark: language-model use is still dominated by retrieval and text question answering, while agents acting on power-system artifacts remain mostly academic prototypes. We introduce the Power Systems Agent Benchmark, an executable benchmark for power-engineering agents. An agent receives a structured task and returns a structured solution; a deterministic evaluator recomputes the engineering quantities, checks operational constraints, and returns a feasibility flag, a normalized score, and explicit violations. The benchmark contains 41 task families across eight areas of power engineering, from power flow and protection to stability, microgrids, reliability, power quality, and forecasting. Each task is grounded in a citable source, standard, or documented engineering formulation. To resist contamination, held-out cases are synthesized on demand by per-family generators from private seeds: the construction is inspectable, but the instances remain private. In a reference evaluation with three command-line agents, the strongest score near the compact tier's ceiling, a smaller open model trails, and public and held-out performance are broadly consistent; a separate public-split grid with OpenCode and Aider probes harness effects. The reference evaluation doubles as quality control: unanimous failures flag candidate task or evaluator defects, and it exposed a latent evaluator bug missed by self-consistency checks. The evaluators are compact deterministic surrogates, but the task contract allows their internals to be upgraded to simulator-backed checks without changing how tasks are posed or solved.

LLM coding agents are increasingly deployed to act autonomously on real production infrastructure. They execute shell commands, modify repositories, and call operational APIs. However, completing a task is not sufficient for safety. A wrong action can cause severe consequences. Existing agent benchmarks largely emphasize task completion, leaving open how agents behave under benign but underspecified instructions. We present UnderSpecBench, a benchmark for measuring action-boundary violations in coding agents (i.e., Claude Code, Codex, and OpenCode) on DevOps tasks. UnderSpecBench includes 69 task families grounded in documented incidents, CVEs, or tool behavior and organized across four DevOps capability domains and nine operational control surfaces. To isolate underspecification from task difficulty, each task keeps the same environment and ground-truth safe action while varying the instruction along three axes: intent clarity, target certainty, and blast radius. The resulting 2,208 prompt variants are evaluated with deterministic, side-effect-based oracles that separate Safe Success, Wrong Target, and OverScope outcomes; non-action runs are further classified as clarification, refusal, or deferment. Across five agent x model configurations using OpenCode, Claude Code, and Codex, the evaluation results show that underspecification does not mainly make agents fail; it makes them guess. 55.8-67.8% of runs violate at least one boundary. Target underspecification sharply degrades action quality, while blast-radius cues barely reduce action propensity. These findings show that completion-centric evaluation can overstate safe autonomy and motivate mitigations at the model, harness, and system layer.

Computed tomography (CT) is a central to three-dimensional medical imaging, yet CT-based artificial intelligence remains fragmented across task-specific models for segmentation, classification, registration, and report analysis. Here we present FlexiCT, a family of CT foundation models trained by agglomerative continual pretraining on 266,227 CT volumes from 56 publicly available datasets, forming a large-scale public resource for CT representation learning. FlexiCT uses agglomerative pretraining across three stages: two-dimensional axial pretraining, three-dimensional anatomical pretraining and report-guided semantic alignment. This training strategy supports slice-level, volume-level and vision-language analysis. Across five downstream task families (segmentation, classification, registration, vision-language understanding and clinical retrieval), FlexiCT matches or exceeds prior task-specific approaches on multiple benchmarks. Its embeddings further organize CT scans along gradients associated with various tumor stages, suggesting that CT foundation models can capture imaging features relevant to disease phenotype characterization. Code is available at https://github.com/ricklisz/FlexiCT

Understanding scientific papers requires more than answering isolated questions or summarizing content. It involves an integrated reasoning process that grounds textual and visual information, interprets experimental evidence, synthesizes information across sources, and critically evaluates scientific claims. However, existing benchmarks typically assess these abilities in isolation, making it difficult to evaluate scientific paper understanding as a unified set of interacting cognitive abilities. In this work, we introduce PAPERMIND, a benchmark designed to evaluate integrated and agent-oriented scientific reasoning over research papers. PAPERMIND is constructed from real scientific papers across seven domains, including agriculture, biology, chemistry, computer science, medicine, physics, and economics. It comprises four complementary task families that collectively operationalize distinct cognitive facets of scientific paper reasoning, including multimodal grounding, experimental interpretation, cross-source evidence reasoning, and critical assessment. By analyzing model behavior across multiple tasks, PAPERMIND enables a diagnostic evaluation of integrated scientific reasoning behaviors that are difficult to assess through isolated task evaluations. Extensive experiments on both opensource and closed-source multimodal LLMs reveal consistent performance gaps across tasks, highlighting persistent challenges in integrated scientific reasoning and critique. Our benchmark and dataset are available at https:// github.com/Yanjun-Zhao/PaperMind.

Generalizable manipulation skills, which can be composed to tackle long-horizon and complex daily chores, are one of the cornerstones of Embodied AI. However, existing benchmarks, mostly composed of a suite of simulatable environments, are insufficient to push cutting-edge research works because they lack object-level topological and geometric variations, are not based on fully dynamic simulation, or are short of native support for multiple types of manipulation tasks. To this end, we present ManiSkill2, the next generation of the SAPIEN ManiSkill benchmark, to address critical pain points often encountered by researchers when using benchmarks for generalizable manipulation skills. ManiSkill2 includes 20 manipulation task families with 2000+ object models and 4M+ demonstration frames, which cover stationary/mobile-base, single/dual-arm, and rigid/soft-body manipulation tasks with 2D/3D-input data simulated by fully dynamic engines. It defines a unified interface and evaluation protocol to support a wide range of algorithms (e.g., classic sense-plan-act, RL, IL), visual observations (point cloud, RGBD), and controllers (e.g., action type and parameterization). Moreover, it empowers fast visual input learning algorithms so that a CNN-based policy can collect samples at about 2000 FPS with 1 GPU and 16 processes on a regular workstation. It implements a render server infrastructure to allow sharing rendering resources across all environments, thereby significantly reducing memory usage. We open-source all codes of our benchmark (simulator, environments, and baselines) and host an online challenge open to interdisciplinary researchers.

Long-context reasoning is an essential capability for large language models, particularly when they are deployed as autonomous agents that must reason over lengthy trajectories. Reinforcement learning (RL) has recently emerged as a dominant paradigm for improving this ability, yet existing work largely focuses on reward engineering while diverse training data remains scarce. We revisit this problem from a data-centric perspective and show that a simple yet effective data recipe alone, paired with a minimal outcome-based GRPO setup, suffices to substantially improve long-context reasoning. Our recipe targets three complementary task families -- retrieval, multi-evidence synthesis, and reasoning -- for which we construct and curate eight datasets totaling ~14K examples. Experiments on three models (Qwen3-4B/8B/30B-A3B) yield average gains of +7.2/+3.2/+6.4 points across seven long-context benchmarks, surpassing prior RL training sets. We further demonstrate that these gains transfer to agentic tasks, where continuing RL training on an agent-tuned model with our data recipe improves GAIA by +4.8 and BrowseComp by +7.0 points. We will release our datasets to facilitate future research.

Evaluation of robotic manipulation systems has largely relied on fixed benchmarks authored by a small number of experts, where task instances, constraints, and success criteria are predefined and difficult to extend. This paradigm limits who can shape evaluation and obscures how policies respond to user-authored variations in task intent, constraints, and notions of success. We argue that evaluating modern manipulation policies requires reframing evaluation as a language-driven process over structured physical domains. We present RoboPlayground, a framework that enables users to author executable manipulation tasks using natural language within a structured physical domain. Natural language instructions are compiled into reproducible task specifications with explicit asset definitions, initialization distributions, and success predicates. Each instruction defines a structured family of related tasks, enabling controlled semantic and behavioral variation while preserving executability and comparability. We instantiate RoboPlayground in a structured block manipulation domain and evaluate it along three axes. A user study shows that the language-driven interface is easier to use and imposes lower cognitive workload than programming-based and code-assist baselines. Evaluating learned policies on language-defined task families reveals generalization failures that are not apparent under fixed benchmark evaluations. Finally, we show that task diversity scales with contributor diversity rather than task count alone, enabling evaluation spaces to grow continuously through crowd-authored contributions. Project Page: https://roboplayground.github.io

In classical Reinforcement Learning from Human Feedback (RLHF), Reward Models (RMs) serve as the fundamental signal provider for model alignment. As Large Language Models evolve into agentic systems capable of autonomous tool invocation and complex reasoning, the paradigm of reward modeling faces unprecedented challenges--most notably, the lack of benchmarks specifically designed to assess RM capabilities within tool-integrated environments. To address this gap, we present Plan-RewardBench, a trajectory-level preference benchmark designed to evaluate how well judges distinguish preferred versus distractor agent trajectories in complex tool-using scenarios. Plan-RewardBench covers four representative task families -- (i) Safety Refusal, (ii) Tool-Irrelevance / Unavailability, (iii) Complex Planning, and (iv) Robust Error Recovery -- comprising validated positive trajectories and confusable hard negatives constructed via multi-model natural rollouts, rule-based perturbations, and minimal-edit LLM perturbations. We benchmark representative RMs (generative, discriminative, and LLM-as-Judge) under a unified pairwise protocol, reporting accuracy trends across varying trajectory lengths and task categories. Furthermore, we provide diagnostic analyses of prevalent failure modes. Our results reveal that all three evaluator families face substantial challenges, with performance degrading sharply on long-horizon trajectories, underscoring the necessity for specialized training in agentic, trajectory-level reward modeling. Ultimately, Plan-RewardBench aims to serve as both a practical evaluation suite and a reusable blueprint for constructing agentic planning preference data.

The success of many RL techniques heavily relies on human-engineered dense rewards, which typically demand substantial domain expertise and extensive trial and error. In our work, we propose DrS (Dense reward learning from Stages), a novel approach for learning reusable dense rewards for multi-stage tasks in a data-driven manner. By leveraging the stage structures of the task, DrS learns a high-quality dense reward from sparse rewards and demonstrations if given. The learned rewards can be reused in unseen tasks, thus reducing the human effort for reward engineering. Extensive experiments on three physical robot manipulation task families with 1000+ task variants demonstrate that our learned rewards can be reused in unseen tasks, resulting in improved performance and sample efficiency of RL algorithms. The learned rewards even achieve comparable performance to human-engineered rewards on some tasks. See our project page (https://sites.google.com/view/iclr24drs) for more details.

In-context learning is a new learning paradigm where a language model conditions on a few input-output pairs (demonstrations) and a test input, and directly outputs the prediction. It has been shown highly dependent on the provided demonstrations and thus promotes the research of demonstration retrieval: given a test input, relevant examples are retrieved from the training set to serve as informative demonstrations for in-context learning. While previous works focus on training task-specific retrievers for several tasks separately, these methods are often hard to transfer and scale on various tasks, and separately trained retrievers incur a lot of parameter storage and deployment cost. In this paper, we propose Unified Demonstration Retriever (UDR), a single model to retrieve demonstrations for a wide range of tasks. To train UDR, we cast various tasks' training signals into a unified list-wise ranking formulation by language model's feedback. Then we propose a multi-task list-wise ranking training framework, with an iterative mining strategy to find high-quality candidates, which can help UDR fully incorporate various tasks' signals. Experiments on 30+ tasks across 13 task families and multiple data domains show that UDR significantly outperforms baselines. Further analyses show the effectiveness of each proposed component and UDR's strong ability in various scenarios including different LMs (1.3B - 175B), unseen datasets, varying demonstration quantities, etc.

As large language models and harness frameworks continue to advance, agents operating in terminals are increasingly capable of performing a broader range of general computer-use tasks beyond coding. However, existing benchmarks do not adequately evaluate general-purpose terminal computer-use agents (TUAs): general computer-use benchmarks primarily target graphical user interfaces (GUIs), whereas terminal-based benchmarks largely emphasize technical and programming-centric workflows historically native to the shell. We introduce TUA-Bench, a general-purpose benchmark for terminal-use agents. TUA-Bench includes 120 real-world tasks across five task families, covering routine digital activities-including document editing, email management, and live-web information seeking-as well as scientific and engineering workflows co-designed with PhD-level domain experts that require specialized software. This breadth distinguishes TUA-Bench from prior shell-focused or domain-specific benchmarks. Each task is manually designed, runs in a real terminal with a deterministic setup script, and is evaluated by an execution-based scoring protocol. We find that the strongest frontier agent, Claude Code with Claude Opus 4.8 max reasoning effort, achieves 65.8% overall performance, with substantial gaps across both tracks. By providing a broad and realistic evaluation of terminal-use capabilities, TUA-Bench aims to accelerate the transition from narrow, task-specific assistants to general-purpose agents capable of operating reliably across diverse digital environments.

Reinforcement Learning with Verifiable Rewards algorithms like GRPO have emerged as the dominant post-training paradigm for complex reasoning in LLMs, yet commonly suffer from policy entropy collapse during training. We conduct a first-order gradient analysis of token-level entropy dynamics under GRPO and identify a token-level credit assignment mismatch: the per-token entropy variation decomposes into the product of the trajectory-level advantage and an entropy sensitivity function over the next-token distribution, yielding an advantage-surprisal four-quadrant structure and a near-criticality property. Motivated by it, we propose STARE (Surprisal-guided Token-level Advantage Reweighting for policy Entropy stability), which identifies entropy-critical token subsets via batch-internal surprisal quantiles, selectively reweights their effective advantages, and incorporates a target-entropy closed-loop gate for stable entropy regulation. Across model scales from 1.5B to 32B and three task families (Short CoT, Long CoT, and Multi-Turn Tool Use), STARE sustains stable RL training over thousands of steps while maintaining policy entropy within the target band. On AIME24 and AIME25, STARE outperforms DAPO and other competitive baselines by 4%-8% in average accuracy, with reflection tokens and response length growing in tandem, indicating sustained exploration-exploitation balance that further unlocks RL training potential.Code is available at https://github.com/hp-luo/STARE.

Omni-modal large language models (OLLMs) aim to unify audio, vision, and text understanding within a single framework. While existing benchmarks primarily evaluate general cross-modal question-answering ability, it remains unclear whether OLLMs achieve modality-invariant reasoning or exhibit modality-specific biases. We introduce XModBench, a large-scale tri-modal benchmark explicitly designed to measure cross-modal consistency. XModBench comprises 60,828 multiple-choice questions spanning five task families and systematically covers all six modality compositions in question-answer pairs, enabling fine-grained diagnosis of an OLLM's modality-invariant reasoning, modality disparity, and directional imbalance. Experiments show that even the strongest model, Gemini 2.5 Pro, (i) struggles with spatial and temporal reasoning, achieving less than 60% accuracy, (ii) reveals persistent modality disparities, with performance dropping substantially when the same semantic content is conveyed through audio rather than text, and (iii) shows systematic directional imbalance, exhibiting lower consistency when vision serves as context compared to text. These findings indicate that current OLLMs remain far from truly modality-invariant reasoning and position XModBench as a fundamental diagnostic tool for evaluating and improving cross-modal competence. All data and evaluation tools will be available at https://xingruiwang.github.io/projects/XModBench/.

Portuguese remains underrepresented in text embedding evaluation, despite being one of the most widely spoken languages in the world. As a result, embedding models are often selected based on English or multilingual metrics, while their effectiveness in Portuguese remains unclear. We present MTEB-PT, a Portuguese benchmark constructed from a subset of MMTEB, comprising 14 existing datasets across Semantic Textual Similarity (STS), classification, retrieval, and reranking. We use this benchmark to evaluate 17 open- and closed-source embedding models under a unified protocol. Our results show that Portuguese performance is strongly task-dependent: multilingual rankings do not reliably predict Portuguese-specific performance across task families, no single model dominates all settings, and models with stronger long-context capacity are particularly advantageous on longer-input tasks such as retrieval and reranking. The benchmark also shows that language-specific fine-tuning still improves model performance in Portuguese, especially on task types that match the adaptation data most closely. To examine this effect, we fine-tune three representative backbone models with Portuguese contrastive supervision and Matryoshka Representation Learning (MRL). These benchmark-informed baselines yield their strongest gains on STS, consistent with the predominantly symmetric supervision used during training, while also improving retrieval and remaining competitive under dimensional truncation. We release the MTEB-PT benchmark, the fine-tuned models, and the training and evaluation code.

Vision-language models (VLMs) are increasingly explored as visual critics, reward generators, and failure detectors in robotic manipulation. These roles implicitly require models to judge not only final task success, but also how a manipulation execution is physically and temporally progressing. However, existing evaluations fail to test whether VLMs possess fine-grained process understanding. To address this gap, we present RoboProcessBench, a benchmark for process-aware understanding in vision-language robotic manipulation. RoboProcessBench decomposes such capability into two complementary dimensions, static monitoring and dynamic reasoning, instantiated as 12 diagnostic question families covering phase, contact, motion, coordination, primitive-local progress, temporal order, outcome, and primitive-level transitions. Built from physically grounded execution traces, the curated benchmark corpus ProcessData contains \textasciitilde 58k question-answer pairs across 260 manipulation tasks, which is further split into ProcessData-SFT and ProcessData-Eval for post-training and evaluation purposes. Extensive evaluation of various VLMs on ProcessData-Eval reveals broad limitations across 12 diagnostic task families, suggesting current models still lack robust process-aware understanding of manipulation executions. But with ProcessData-SFT, the post-trained Qwen2.5-VL-7B and InternVL-3-8B exhibit consistent gains on local state, motion, progress, and primitive-aware cues. These results demonstrate that RoboProcessBench serves as both an evaluation benchmark and a learnable supervision source for developing VLMs capable of monitoring and evaluating robotic manipulation processes. Project webpage: https://processbench-2026.github.io/RoboProcessBench-Web/{https://processbench-2026.github.io}.

Supramolecular chemistry, which includes the study of non-covalent host-guest assemblies, has advanced various applications. However, designing host-guest systems remains time-consuming, requiring days of dry-lab verification per candidate pair. Although LLMs have emerged as a fast alternative with strong performance on molecular binding tasks, no benchmark currently systematically evaluates LLMs for host-guest reasoning across fundamental supramolecular chemistry tasks, e.g., binding affinity prediction. To this end, we collaborate with domain experts to release the first Supramolecular Benchmark, called SupraBench, to evaluate LLMs in chemistry reasoning. Specifically, we design four fundamental tasks, i.e., binding affinity prediction, top-binder selection, solvent identification, and host-guest description, plus an auxiliary vision-based task for molecular identification. We also release SupraPMC, a curated 16M-token corpus of Supramolecular chemistry articles distilled from Europe PMC, to support the adaptation to the supramolecular domain. We benchmark a broad range of open and proprietary LLMs and find that LLMs leave substantial headroom across all tasks. Domain adaptation pretraining over SupraPMC transfers cleanly to in-distribution regression but trades off against strict letter-format output. Moreover, the difficulty profile differs sharply across task families, revealing distinct failure modes that indicate specific gaps in current supramolecular chemistry reasoning. Our source codes and benchmark datasets are available at https://github.com/Tianyi-Billy-Ma/SupraBench.

Patent text embeddings enable prior art search, technology landscaping, and patent analysis, yet existing benchmarks inadequately capture patent-specific challenges. We introduce PatenTEB, a comprehensive benchmark comprising 15 tasks across retrieval, classification, paraphrase, and clustering, with 2.06 million examples. PatenTEB employs domain-stratified splits, domain specific hard negative mining, and systematic coverage of asymmetric fragment-to-document matching scenarios absent from general embedding benchmarks. We develop the patembed model family through multi-task training, spanning 67M to 344M parameters with context lengths up to 4096 tokens. External validation shows strong generalization: patembed-base achieves state-of-the-art on MTEB BigPatentClustering.v2 (0.494 V-measure vs. 0.445 previous best), while patembed-large achieves 0.377 NDCG@100 on DAPFAM. Systematic ablations reveal that multi-task training improves external generalization despite minor benchmark costs, and that domain-pretrained initialization provides consistent advantages across task families. All resources will be made available at https://github.com/iliass-y/patenteb. Keywords: patent retrieval, sentence embeddings, multi-task learning, asymmetric retrieval, benchmark evaluation, contrastive learning.

Reinforcement learning (RL) trained language model agents with tool access are increasingly deployed in coding assistants, research tools, and autonomous systems. We introduce the Reward Hacking Benchmark (RHB), a suite of multi-step tasks requiring sequential tool operations with naturalistic shortcut opportunities such as skipping verification steps, inferring answers from task-adjacent metadata, or tampering with evaluation-relevant functions. RHB supports independent and chained task regimes, where chain length acts as a proxy for longer-horizon agent behavior. We evaluate 13 frontier models from OpenAI, Anthropic, Google, and DeepSeek. Exploit rates range from 0% (Claude Sonnet 4.5) to 13.9% (DeepSeek-R1-Zero), varying sharply by post-training style. A controlled sibling comparison (DeepSeek-V3 vs. DeepSeek-R1-Zero) shows RL post-training is associated with substantially higher reward hacking (0.6% vs. 13.9%), with consistent gaps across all four task families. We identify six exploit categories and find that 72% of reward hacking episodes include explicit chain-of-thought rationale, suggesting models often frame exploits as legitimate problem-solving. Simple environmental hardening reduces exploit rates by 5.7 percentage points (87.7% relative) without degrading task success. Models with near-zero exploit rates on standard tasks show elevated rates on harder variants, suggesting that production-aligned post-training appears to suppress reward hacking only below a complexity threshold where honest solutions remain tractable.

Multimodal large language models can write code to produce complex programs as well as use programs to do 3D modeling, which opens up a new avenue for 3D generation powered by their priors, world knowledge and reasoning. Yet existing benchmarks rarely evaluate 3D modeling through code. Such modeling demands more than runnable code: from a text or visual specification, a model must generate a parametric 3D program that is geometrically precise, semantically aligned and assembly-consistent. We introduce P3D-Bench, a benchmark for parametric 3D generation. Unlike a 3D mesh, a parametric 3D program exposes explicit dimensions, construction operations and part relations, revealing whether a model recovers a design's structure, not just its appearance. Under a unified protocol, P3D-Bench covers three task families (Text-to-3D, Image-to-3D and Assembly-3D) and scores each output for executability, geometric fidelity, topology, text-grounded constraints, multiview semantic alignment and part-level structure. We evaluate frontier MLLMs and text-only LLMs on 400 text cases, 400 image cases and 203 annotated assemblies, with domain-specific models as reference points. Our extensive evaluation yields three findings. First, assemblies are the hardest setting, where models still fail to compose multiple parts into a coherent structure. Second, models can often recover the global shape and semantic identity of the target object, yet fail to reproduce the precise parametric geometry specified by the input. Third, part-level modeling remains weak on assemblies, where models recover neither the geometry of each part nor the right number of parts. These results position P3D-Bench as a benchmark for evaluating precise parametric geometry and part-level structure in parametric 3D generation.

Large language model (LLM) agents accumulate rich episodic trajectories while solving real-world tasks, but it remains unclear whether such experience can be distilled into reusable procedural skills. We introduce SkillEvolBench, a diagnostic benchmark for evaluating this step from experience reuse to skill formation. It contains 180 tasks across six real-world agent environments, organized into role-conditioned task families with shared latent procedures. Agents learn from acquisition tasks, update an external skill library using compacted trajectories and verifier feedback, and then face frozen deployment tasks testing context shift, adversarial shortcuts, and composition. By comparing self-generated and curated-start skill evolution against no-skill and raw-trajectory controls, SkillEvolBench separates procedural abstraction from base capability, curated prior knowledge, and direct reuse of episodic traces. Across ten model configurations and three agent harnesses, we find that current agents often adapt locally but rarely form robust reusable skills. Skill-based conditions can improve acquisition or replay, and individual models sometimes gain on specific deployment axes, but these gains are unstable under frozen deployment. Raw-trajectory reuse frequently outperforms distilled skills, suggesting that current abstraction procedures discard contextual and procedural cues that remain useful for future tasks. Capacity and cost analyses further show that writing more skills or larger Tier-3 resource libraries is not sufficient: additional updates can improve coverage while introducing episode-specific drift and procedural clutter. These findings position SkillEvolBench as a testbed for measuring when one-off experience becomes durable procedural knowledge rather than task-local memory.

Speech language models (SpeechLMs) have achieved substantial progress by extending large language models (LLMs) to the speech modality. However, SpeechLM evaluation remains heavily centered on English, limiting reliable assessment of multilingual speech capabilities. Straightforward benchmark transfer through ASR, translation, normalization, and TTS can corrupt language-specific instructions, answer constraints, and spoken forms; for audio understanding, transferring source-language audio also fails to preserve target-language speaker attributes, accents, and paralinguistic properties. To address these limitations, we propose two human-agent benchmark-construction frameworks: one transfers source-language SpokenQA benchmarks into target-language SpokenQA benchmarks, and the other converts target-language ASR corpora into audio understanding benchmarks using transcriptions and speaker metadata. Using these frameworks, we construct and publicly release three Korean speech benchmarks: KVoiceBench and KOpenAudioBench for Korean SpokenQA, and KMMAU for Korean audio understanding, comprising 12,345 samples in total. We evaluate eight recent SpeechLMs and find that English-Korean performance gaps vary substantially across models and task families, and that SpokenQA and audio understanding rankings diverge, revealing complementary weaknesses invisible to English-only evaluation.

Longitudinal chest X-ray (CXR) interpretation requires reasoning over disease evolution across multiple patient visits, yet most existing medical VQA benchmarks focus on single images or short-horizon image pairs. We introduce MI-CXR, a benchmark for standardized evaluation of Multi-Interval longitudinal reasoning over multi-visit CXR sequences, without requiring free-form report generation or additional clinical context. MI-CXR comprises five-way multiple-choice questions over five-visit patient timelines and instantiates three complementary task families: Temporal Event Localization, Interval-wise Change Reasoning, and Global Trajectory Summarization, which assess clinically grounded visual reasoning over time. Evaluating 14 state-of-the-art vision-language models (VLMs) shows low overall performance, with an average accuracy of 29.3%, only modestly above random guessing. Using stage-wise diagnostic probing, we find that models often produce locally plausible interval descriptions but fail to enforce temporal constraints or compose evidence into globally consistent decisions over the full timeline. These findings reveal key limitations of current VLMs and establish MI-CXR as a principled benchmark for longitudinal medical reasoning. The benchmark is available at https://github.com/AIDASLab/MI-CXR

Agentic navigation systems require a base navigation model whose observation strategy can be externally reconfigured at inference time, because instruction following, object search, target tracking, and autonomous driving share the same perception-planning backbone yet demand fundamentally different strategies for consuming the visual stream. We present Qwen-RobotNav, a scalable navigation model built on Qwen-RobotNav that addresses it through a parameterised interface with two complementary dimensions: multiple task modes that select the navigation behaviour, and controllable observation parameters (e.g., token budget, per-camera weights) that govern how visual history is encoded. With training-time randomization over all parameters, Qwen-RobotNav is robust to any inference-time configuration requiring zero architectural modification to the Qwen-RobotNav backbone. We train Qwen-RobotNav on 15.6M samples; co-training with vision-language data prevents the collapse into reactive action-sequence mappers observed in trajectory-only training. The parameterised interface also makes Qwen-RobotNav a natural building block for agentic systems: for long-horizon scenarios, an upper-level planner decomposes goals into sub-tasks and dynamically switches Qwen-RobotNav's task mode and context strategy mid-episode, composing complex behaviours from repeated calls to the same model. Extensive experiments show that Qwen-RobotNav sets new state-of-the-art results across major navigation benchmarks. The model exhibits favourable scaling from 2B to 8B parameters, with joint multi-task training developing a shared spatial-planning substrate that transfers across task families, and demonstrates strong zero-shot generalisation to real-world robots across diverse environments.

Humans are remarkably efficient at learning tasks from demonstrations, but today's imitation learning methods for robot manipulation often require hundreds or thousands of demonstrations per task. We investigate two fundamental priors for improving learning efficiency: decomposing manipulation trajectories into sequential alignment and interaction phases, and retrieval-based generalisation. Through 3,450 real-world rollouts, we systematically study this decomposition. We compare different design choices for the alignment and interaction phases, and examine generalisation and scaling trends relative to today's dominant paradigm of behavioural cloning with a single-phase monolithic policy. In the few-demonstrations-per-task regime (<10 demonstrations), decomposition achieves an order of magnitude improvement in data efficiency over single-phase learning, with retrieval consistently outperforming behavioural cloning for both alignment and interaction. Building on these insights, we develop Multi-Task Trajectory Transfer (MT3), an imitation learning method based on decomposition and retrieval. MT3 learns everyday manipulation tasks from as little as a single demonstration each, whilst also generalising to novel object instances. This efficiency enables us to teach a robot 1,000 distinct everyday tasks in under 24 hours of human demonstrator time. Through 2,200 additional real-world rollouts, we reveal MT3's capabilities and limitations across different task families. Videos of our experiments can be found on at https://www.robot-learning.uk/learning-1000-tasks.

Large language models are increasingly used in K-12 education, but existing benchmarks mainly test exam question answering rather than understanding how curriculum knowledge is structured and visually presented. We call this capability curriculum cognition. It covers prerequisite chains, concept taxonomies, experiment-concept links, pedagogical sequencing, and visual grounding. We introduce K12-KGraph, a curriculum-aligned knowledge graph extracted from official People's Education Press textbooks in mathematics, physics, chemistry, and biology across primary, middle, and high school. It contains nine node types and fourteen relation types covering curriculum structure and visual grounding. From this graph, we derive K12-Bench, a 23,640-question multi-select benchmark with five task families: Ground, Prereq, Neighbor, Evidence, and Locate. We also build K12-Train, a graph-guided supervised fine-tuning corpus of 7,335 samples, including 2,267 text-only QA pairs and 5,068 multimodal VQA pairs. On K12-Bench, Gemini-3-Flash achieves only 57 percent exact match and Gemma-4-31B-IT reaches 46 percent, with Prereq and Neighbor being the hardest tasks. Our training experiments show that domain-specific supervision can reduce this gap. Under a matched 2,300-sample budget, K12-Train-Text consistently outperforms equally sized subsets of eight mainstream instruction-tuning corpora on GaokaoBench and EduEval. For vision-language models, K12-Train-Full achieves the best overall results on Gaokao-MM, MDK12-medium, and K12Vista among all compared training configurations, despite using fewer samples than the full DataFlow and WizardLM baselines. It also surpasses both text-only and multimodal-only variants, showing that textual and visual supervision are complementary. We release the graph, benchmark, training data, and complete construction pipeline.

Audio-Visual Intelligence (AVI) has emerged as a central frontier in artificial intelligence, bridging auditory and visual modalities to enable machines that can perceive, generate, and interact in the multimodal real world. In the era of large foundation models, joint modeling of audio and vision has become increasingly crucial, i.e., not only for understanding but also for controllable generation and reasoning across dynamic, temporally grounded signals. Recent advances, such as Meta MovieGen and Google Veo-3, highlight the growing industrial and academic focus on unified audio-vision architectures that learn from massive multimodal data. However, despite rapid progress, the literature remains fragmented, spanning diverse tasks, inconsistent taxonomies, and heterogeneous evaluation practices that impede systematic comparison and knowledge integration. This survey provides the first comprehensive review of AVI through the lens of large foundation models. We establish a unified taxonomy covering the broad landscape of AVI tasks, ranging from understanding (e.g., speech recognition, sound localization) to generation (e.g., audio-driven video synthesis, video-to-audio) and interaction (e.g., dialogue, embodied, or agentic interfaces). We synthesize methodological foundations, including modality tokenization, cross-modal fusion, autoregressive and diffusion-based generation, large-scale pretraining, instruction alignment, and preference optimization. Furthermore, we curate representative datasets, benchmarks, and evaluation metrics, offering a structured comparison across task families and identifying open challenges in synchronization, spatial reasoning, controllability, and safety. By consolidating this rapidly expanding field into a coherent framework, this survey aims to serve as a foundational reference for future research on large-scale AVI.

Reasoning has become a core capability for large models, especially when reliable decisions require understanding logical consequences. Recent video generation models offer a reasoning path distinct from previous Chain-of-Thought (CoT): reasoning can unfold through temporally connected frames, known as Chain-of-Frame (CoF) reasoning. However, existing video generators are primarily trained on general video corpora, still lacking diverse supervision and dedicated designs for CoF reasoning. To address this gap, we introduce OpenCoF, a framework comprising the OpenCoF-17K dataset, a reasoning video dataset spanning 11 task families, and Wan-CoF, a fine-tuned video model for studying whether diverse temporal supervision improves CoF behavior. Across four video reasoning benchmarks, Wan-CoF achieves considerable gains over the Wan2.2-I2V-A14B baseline. Building on this, we empirically explore more advanced designs for CoF capabilities, i.e., equipping the model with visual and textual reasoning tokens. This mechanism respectively captures low-level visual cues and high-level semantic priors for spatial and temporal reasoning. Through performance comparisons and attention analysis, we examine how these tokens contribute across model depth, denoising steps, space, and time. Our results suggest that stronger video reasoning requires both broad temporal supervision and explicit mechanisms for organizing intermediate reasoning state. We open-source the dataset, model, and code to facilitate future research on reasoning-oriented video generation.

The ability to predict future outcomes given control actions is fundamental for physical reasoning. However, such predictive models, often called world models, have proven challenging to learn and are typically developed for task-specific solutions with online policy learning. We argue that the true potential of world models lies in their ability to reason and plan across diverse problems using only passive data. Concretely, we require world models to have the following three properties: 1) be trainable on offline, pre-collected trajectories, 2) support test-time behavior optimization, and 3) facilitate task-agnostic reasoning. To realize this, we present DINO World Model (DINO-WM), a new method to model visual dynamics without reconstructing the visual world. DINO-WM leverages spatial patch features pre-trained with DINOv2, enabling it to learn from offline behavioral trajectories by predicting future patch features. This design allows DINO-WM to achieve observational goals through action sequence optimization, facilitating task-agnostic behavior planning by treating desired goal patch features as prediction targets. We evaluate DINO-WM across various domains, including maze navigation, tabletop pushing, and particle manipulation. Our experiments demonstrate that DINO-WM can generate zero-shot behavioral solutions at test time without relying on expert demonstrations, reward modeling, or pre-learned inverse models. Notably, DINO-WM exhibits strong generalization capabilities compared to prior state-of-the-art work, adapting to diverse task families such as arbitrarily configured mazes, push manipulation with varied object shapes, and multi-particle scenarios.

Instruction-based image editing has emerged as a key capability for unified multimodal models (UMMs), yet constructing large-scale, diverse, and high-quality editing datasets without costly proprietary APIs remains challenging. Previous image editing datasets either rely on closed-source models for annotation, which prevents cost-effective scaling, or employ fixed synthetic editing pipelines, which suffer from limited quality and generalizability. To address these challenges, we propose ScaleEditor, a fully open-source hierarchical multi-agent framework for end-to-end construction of large-scale, high-quality image editing datasets. Our pipeline consists of three key components: source image expansion with world-knowledge infusion, adaptive multi-agent editing instruction-image synthesis, and a task-aware data quality verification mechanism. Using ScaleEditor, we curate ScaleEdit-12M, the largest open-source image editing dataset to date, spanning 23 task families across diverse real and synthetic domains. Fine-tuning UniWorld-V1 and Bagel on ScaleEdit yields consistent gains, improving performance by up to 10.4% on ImgEdit and 35.1% on GEdit for general editing benchmarks and by up to 150.0% on RISE and 26.5% on KRIS-Bench for knowledge-infused benchmarks. These results demonstrate that open-source, agentic pipelines can approach commercial-grade data quality while retaining cost-effectiveness and scalability. Both the framework and dataset will be open-sourced.

Can Large Language Models understand how students learn? As LLMs are deployed for adaptive testing and personalized tutoring, this question becomes urgent -- yet we cannot answer it with existing resources. Current educational datasets provide only question identifiers and binary correctness labels, rendering them opaque to LLMs that reason in natural language. We address this gap with FoundationalASSIST, the first English educational dataset providing the complete information needed for research on LLMs in education: full question text, actual student responses (not just right/wrong), records of which wrong answers students chose, and alignment to Common Core K-12 standards. These 1.7 million interactions from 5,000 students enable research directions that were previously impossible to pursue, from fine-tuning student models to analyzing misconception patterns. To demonstrate the dataset's utility, we evaluate four frontier models (GPT-OSS-120B, Llama-3.3-70B, Qwen3-Next-80B variants) on two complementary task families: Knowledge Tracing, testing whether LLMs can predict student performance on questions, and the exact answer a student will give; and Pedagogical Grounding, testing whether LLMs understand the properties that make assessment items effective. Our evaluation reveals significant gaps in current LLM capabilities. Every model barely achieves a trivial baseline on knowledge tracing. All models fall below random chance on item discrimination, indicating that LLMs do not understand what makes one problem more diagnostic than another. Models do show competence at judging relative difficulty (up to 68.6%), but this partial success only highlights the gaps elsewhere. These results establish that substantial advances are needed before LLMs can reliably support personalized learning at scale. We release FoundationalASSIST to support progress on these foundational challenges.

Transformer models, notably large language models (LLMs), have the remarkable ability to perform in-context learning (ICL) -- to perform new tasks when prompted with unseen input-output examples without any explicit model training. In this work, we study how effectively transformers can bridge between their pretraining data mixture, comprised of multiple distinct task families, to identify and learn new tasks in-context which are both inside and outside the pretraining distribution. Building on previous work, we investigate this question in a controlled setting, where we study transformer models trained on sequences of (x, f(x)) pairs rather than natural language. Our empirical results show transformers demonstrate near-optimal unsupervised model selection capabilities, in their ability to first in-context identify different task families and in-context learn within them when the task families are well-represented in their pretraining data. However when presented with tasks or functions which are out-of-domain of their pretraining data, we demonstrate various failure modes of transformers and degradation of their generalization for even simple extrapolation tasks. Together our results highlight that the impressive ICL abilities of high-capacity sequence models may be more closely tied to the coverage of their pretraining data mixtures than inductive biases that create fundamental generalization capabilities.

Video production workflows offer a rich and demanding arena for evaluating multimodal AI agents: they require composite capabilities across text, image, audio, and video understanding, along with long-horizon planning, and tool use. To this end, we introduce AgenticVBench, a benchmark of 100 agentic tasks across 4 task families spanning the real world post-production workflow, constructed from real production workflows contributed by 20 industry experts averaging 6 years of professional experience. Tasks are paired with evaluation specifications that combine programmatic verifiers and expert rubrics. We evaluate frontier vision-language models (VLMs) with both vendor-native and open-source harnesses. The best evaluated agent stack barely crosses 30%, far below human expert performance on the same tasks. We further find that the choice of harness substantially affects model behavior, including scores, tool-use patterns, and failure modes. AgenticVBench provides a foundation for diagnosing and improving both models and harnesses for agentic video production. Benchmark website: https://agenticvbench.com.

Product data scientists often ask LLM-based agents to help with recurring execution tasks such as cleaning data, writing SQL, choosing statistical tests, and formatting results. Reusable skill files are meant to avoid prompting from scratch by packaging guidance for a task family. Expert-written skills can encode high-quality guidance, but writing and maintaining them across many data-science task families creates a manual bottleneck. We ask whether LLM-generated skills offer a useful low-curation alternative: do they improve performance over the task prompt alone? We test this question across four lifecycle stages: data preparation, data extraction, statistical analysis, and reporting, using one generated skill per stage. We find no reliable improvement from full generated skills over No-Skill prompting. We then ask whether any part of the skill is useful by ablating different skill components. The main ablation covers 56 tasks, nine model configurations, and three providers, yielding 7,560 runs. Compared with prompting using the task alone, neither the full generated skill nor any ablated skill variant significantly improves performance; all p-values are at least 0.396, and the total spread across variants is only 1.2 pp. A supplemental token-matched control adds 1,512 runs and finds that Full skills perform similarly to task-irrelevant skill-formatted content. The results caution against using one LLM-generated skill per data-science workflow as a default single-shot prompting strategy.

Embodied navigation requires an agent to map language and visual observations to a stream of spatial actions that drive a real robot through environments it has never seen. The dominant approach has been to scale vision-language-action (VLA) foundation models on ever-larger collections of robot trajectories. This paper argues that, for navigation specifically, generality can be obtained structurally, not only through data scale. The underlying decision structure of navigation reduces to a single Language-Vision-Robot Actions Translation. The language action emits semantic-level directional command and the vision action emits a pixel-level visual target. Both outputs lie inside the natural output manifold of pretrained multimodal large language models (MLLMs), so the task can be reasoned about by an agent rather than learned from robot data. Therefore, we present Uni-LaViRA, a unified agentic architecture that extends the same insight to four task families (VLN-CE, ObjectNav, EQA, and Aerial-VLN) and to four heterogeneous real robots (Wheeled, Quadruped, Humanoid robot, and a self-built UAV) in a zero-shot manner. Two agent-loop mechanisms make this unification practical. TODO List Memory (TDM) rewrites a structured checklist of pending sub-goals at every step, reciting the unfinished items back into the agent's most recent attention window. Second Chance Backtrack (SCB) rolls the robot back to the pre-error state and conditions the agent's next plan on the failed sub-trajectory, turning single-pass navigation into a self-correcting process. With zero training effort, Uni-LaViRA reaches 60.7% SR on VLN-CE R2R, 51.3% on VLN-CE RxR, 77.7% on HM3D-v2, 60.0% on HM3D-OVON, 54.7% on MP3D-EQA, and 40.0% on OpenUAV, matching or even surpassing recent training navigation foundation models that consume millions of samples and thousands of GPU-hours.

Vision-Language Models (VLMs) are increasingly deployed in settings where reliable visual grounding carries operational consequences, yet their behavior under progressively coercive prompt phrasing remains undercharacterized. Existing hallucination benchmarks predominantly rely on neutral prompts and binary detection, leaving open how both the incidence and the intensity of fabrication respond to graded linguistic pressure across structurally distinct task types. We present Ghost-100, a procedurally constructed benchmark of 800 synthetically generated images spanning eight categories across three task families: text-illegibility, time-reading, and object-absence, each designed under a negative-ground-truth principle that guarantees the queried target is absent, illegible, or indeterminate by construction. Every image is paired with five prompts drawn from a structured 5-Level Prompt Intensity Framework, holding the image and task identity fixed while varying only directive force, so that tone is isolated as the sole independent variable. We adopt a dual-track evaluation protocol: a rule-based H-Rate measuring the proportion of responses in which a model crosses from grounded refusal into unsupported positive commitment, and a GPT-4o-mini-judged H-Score on a 1-5 scale characterizing the confidence and specificity of fabrication once it occurs. We additionally release a three-stage automated validation workflow, which retrospectively confirms 717 of 800 images as strictly compliant. Evaluating nine open-weight VLMs, we find that H-Rate and H-Score dissociate substantially across model families, reading-style and presence-detection subsets respond to prompt pressure in qualitatively different ways, and several models exhibit non-monotonic sensitivity peaking at intermediate tone levels: patterns that aggregate metrics obscure.

Large Language Models (LLMs) are increasingly adopted as conversational assistants in genomics, where they are mainly used to reason over biological knowledge, annotations, and analysis outputs through natural language interfaces. However, existing benchmarks either focus on specialized DNA models trained for sequence prediction or evaluate biological knowledge using text-only questions, leaving the behavior of general-purpose LLMs when directly exposed to raw genome sequences underexplored. We introduce GenomeQA, a benchmark designed to provide a controlled evaluation setting for general-purpose LLMs on sequence-based genome inference tasks. GenomeQA comprises 5,200 samples drawn from multiple biological databases, with sequence lengths ranging from 6 to 1,000 base pairs (bp), spanning six task families: Enhancer and Promoter Identification, Splice Site Identification, Taxonomic Classification, Histone Mark Prediction, Transcription Factor Binding Site Prediction, and TF Motif Prediction. Across six frontier LLMs, we find that models consistently outperform random baselines and can exploit local sequence signals such as GC content and short motifs, while performance degrades on tasks that require more indirect or multi-step inference over sequence patterns. GenomeQA establishes a diagnostic benchmark for studying and improving the use of general-purpose LLMs on raw genomic sequences.

Recent Multimodal Large Language Models (MLLMs) achieve promising performance on visual and audio benchmarks independently. However, the ability of these models to process cross-modal information synchronously remains largely unexplored. We introduce Daily-Omni, a multiple-choice Audio-Visual QA benchmark featuring 684 real-world videos and 1,197 questions spanning 6 task families that explicitly require cross-modal temporal reasoning. To support scalable benchmark construction, we develop a semi-automatic pipeline for annotation, cross-modal consistency refinement, temporal alignment elicitation, and text-only leakage filtering, followed by human verification. We further provide a diagnostic evaluation suite and extensively evaluate 24 foundation models under 37 model--modality settings (Audio+Video / Audio-only / Video-only / Text-only). Finally, we include a training-free modular diagnostic baseline that composes off-the-shelf unimodal models to serve as a diagnostic baseline and to illustrate how explicit temporal alignment signals affect performance. Results indicate that many end-to-end MLLMs still struggle on alignment-critical questions, suggesting that robust cross-modal temporal alignment remains an important open challenge.

Can neural networks systematically capture discrete, compositional task structure despite their continuous, distributed nature? The impressive capabilities of large-scale neural networks suggest that the answer to this question is yes. However, even for the most capable models, there are still frequent failure cases that raise doubts about their compositionality. Here, we seek to understand what it takes for a standard neural network to generalize over tasks that share compositional structure. We find that simply scaling data and model size leads to compositional generalization. We show that this holds across different task encodings as long as the training distribution sufficiently covers the task space. In line with this finding, we prove that standard multilayer perceptrons can approximate a general class of compositional task families to arbitrary precision using only a linear number of neurons with respect to the number of task modules. Finally, we uncover that if networks successfully compositionally generalize, the constituents of a task can be linearly decoded from their hidden activations. We show that this metric correlates with failures of text-to-image generation models to compose known concepts.

Reinforcement learning (RL) for large language models (LLMs) increasingly relies on sparse, outcome-level rewards -- yet determining which actions within a long trajectory caused the outcome remains difficult. This credit assignment (CA) problem manifests in two regimes: reasoning RL, where credit must be distributed across tokens and steps within a single chain-of-thought generation (500--30K+ tokens); and agentic RL, where multi-turn environment interaction introduces stochastic transitions, partial observability, and horizons of 100+ turns (100K--1M tokens), making episode-level credit increasingly uninformative. We survey 47 CA methods (41 core, 6 adjacent enablers) published between 2024 and early 2026, organizing them in a two-dimensional taxonomy by assignment granularity (token, segment, step, turn, multi-agent) and methodology (Monte Carlo, temporal difference, model-based, game-theoretic, information-theoretic). Beyond the survey itself, we contribute three reusable resources: (1) a structured, machine-readable paper inventory with taxonomy labels, baseline families, and evidence levels; (2) a reporting checklist for future CA papers, validated against the reviewed literature to identify systematic methodological gaps; and (3) a benchmark protocol specification with task families, metadata requirements, and controlled bifurcation tasks, accompanied by a method selection decision tree. Our synthesis suggests that the shift from reasoning to agentic RL complicates and reshapes the credit assignment landscape: reasoning CA is maturing around process reward models and critic-free group comparison, while agentic CA is driving genuinely new approaches -- hindsight counterfactual analysis, privileged asymmetric critics, and turn-level MDP reformulations -- that have no direct precedent in reasoning RL.

Embodied intelligence is often studied through specialized models for individual tasks such as manipulation or navigation, resulting in fragmented capabilities and limited generalization across tasks, environments, and robot embodiments. In this work, we study whether heterogeneous embodied decision-making problems can be unified within a single vision-language-action model. We present Qwen-VLA, a unified embodied foundation model that extends Qwen's vision-language modeling stack from perception, understanding, and reasoning to continuous action and trajectory generation through a DiT-based action decoder. Qwen-VLA is trained with a large-scale joint pretraining recipe over diverse data sources, including robotics manipulation trajectories, human egocentric demonstrations, synthetic simulation data, vision-and-language navigation data, trajectory-centric supervision, and auxiliary vision-language data. To support multiple robot platforms, we introduce embodiment-aware prompt conditioning, where robot-specific textual descriptions specify the current embodiment and control convention. We further cast manipulation, navigation, and trajectory prediction into a unified action-and-trajectory prediction framework, enabling transferable visual grounding, spatial reasoning, and continuous action generation across robot morphologies, task families, and environments. Experiments on manipulation, navigation, and trajectory-centric benchmarks show consistent multi-task performance and out-of-distribution generalization under variations in scene layout, background, lighting, object configuration, and robot embodiment. Qwen-VLA-Instruct achieves 97.9% on LIBERO, 73.7% on Simpler-WidowX, 86.1%/87.2% on RoboTwin-Easy/Hard, 69.0% OSR on R2R, 59.6% SR on RxR, 76.9% average OOD success in real-world ALOHA experiments, and 26.6% zero-shot success on DOMINO dynamic manipulation.

We present the first shared task on Semantic Textual Relatedness (STR). While earlier shared tasks primarily focused on semantic similarity, we instead investigate the broader phenomenon of semantic relatedness across 14 languages: Afrikaans, Algerian Arabic, Amharic, English, Hausa, Hindi, Indonesian, Kinyarwanda, Marathi, Moroccan Arabic, Modern Standard Arabic, Punjabi, Spanish, and Telugu. These languages originate from five distinct language families and are predominantly spoken in Africa and Asia -- regions characterised by the relatively limited availability of NLP resources. Each instance in the datasets is a sentence pair associated with a score that represents the degree of semantic textual relatedness between the two sentences. Participating systems were asked to rank sentence pairs by their closeness in meaning (i.e., their degree of semantic relatedness) in the 14 languages in three main tracks: (a) supervised, (b) unsupervised, and (c) crosslingual. The task attracted 163 participants. We received 70 submissions in total (across all tasks) from 51 different teams, and 38 system description papers. We report on the best-performing systems as well as the most common and the most effective approaches for the three different tracks.

We present our winning system for Task~B (generation with reference passages) in SemEval-2026 Task~8: MTRAGEval. Our method is a heterogeneous ensemble of seven LLMs with two prompting variants, where a GPT-4o-mini judge selects the best candidate per instance. We ranked 1st out of 26 teams, achieving a conditioned harmonic mean of 0.7827 and outperforming the strongest baseline (gpt-oss-120b, 0.6390). Ablations show that diversity in model families, scales, and prompting strategies is essential, with the ensemble consistently beating any single model. We also introduce Meno-Lite-0.1, a 7B domain-adapted model with a strong cost--performance trade-off, and analyse MTRAGEval, highlighting annotation limitations and directions for improvement. Our code is publicly available: https://github.com/RaguTeam/ragu_mtrag_semeval

Financial decision-making is contextual: forecasting prices, valuing companies, and assessing event exposure weigh price history, accounting fundamentals, macroeconomic regime, and contemporaneous text. A benchmark over these four signals is hard to build because finance violates four assumptions of time-series evaluation: text must be gated by its publication date to prevent look-ahead, quarterly fundamentals are reported with a one- to ninety-day lag, filing text is partly redundant with the numerical statement fields it accompanies, and macroeconomic regimes leak across calendar splits. No public benchmark addresses all four signals jointly. MacroLens covers 4,416 U.S. small- and micro-cap equities over 2021-2026. Seven tasks share one point-in-time panel of prices, 46.8M XBRL accounting facts, 53 macroeconomic series, 295,860 SEC filings, and 215,882 news articles, plus a scenario layer of 1,130 macroeconomic events across 49 types automatically detected and rendered as natural language. Tasks span contextual forecasting, public and private valuation, statement generation from fundamentals and descriptions, scenario-conditioned returns, and real-estate valuation. We evaluate 19 methods across six families spanning naive heuristics through time-series foundation models, fine-tuned LLM-based time-series models, and zero-shot large language models (LLMs), plus a five-step feature-context ablation on two frontier LLMs and a gradient-boosted baseline. MacroLens is released at https://huggingface.co/datasets/DeepAuto-AI/MacroLens.

Evaluations of Deep Reinforcement Learning (DRL) methods are an integral part of scientific progress of the field. Beyond designing DRL methods for general intelligence, designing task-specific methods is becoming increasingly prominent for real-world applications. In these settings, the standard evaluation practice involves using a few instances of Markov Decision Processes (MDPs) to represent the task. However, many tasks induce a large family of MDPs owing to variations in the underlying environment, particularly in real-world contexts. For example, in traffic signal control, variations may stem from intersection geometries and traffic flow levels. The select MDP instances may thus inadvertently cause overfitting, lacking the statistical power to draw conclusions about the method's true performance across the family. In this article, we augment DRL evaluations to consider parameterized families of MDPs. We show that in comparison to evaluating DRL methods on select MDP instances, evaluating the MDP family often yields a substantially different relative ranking of methods, casting doubt on what methods should be considered state-of-the-art. We validate this phenomenon in standard control benchmarks and the real-world application of traffic signal control. At the same time, we show that accurately evaluating on an MDP family is nontrivial. Overall, this work identifies new challenges for empirical rigor in reinforcement learning, especially as the outcomes of DRL trickle into downstream decision-making.

The growing collaboration between humans and AI models in generative tasks has introduced new challenges in distinguishing between human-written, LLM-generated, and human--LLM collaborative texts. In this work, we collect a multilingual, multi-domain, multi-generator dataset FAIDSet. We further introduce a fine-grained detection framework FAID to classify text into these three categories, and also to identify the underlying LLM family of the generator. Unlike existing binary classifiers, FAID is built to capture both authorship and model-specific characteristics. Our method combines multi-level contrastive learning with multi-task auxiliary classification to learn subtle stylistic cues. By modeling LLM families as distinct stylistic entities, we incorporate an adaptation to address distributional shifts without retraining for unseen data. Our experimental results demonstrate that FAID outperforms several baselines, particularly enhancing the generalization accuracy on unseen domains and new LLMs, thus offering a potential solution for improving transparency and accountability in AI-assisted writing.

As language models (LMs) become increasingly powerful, it is important to quantify and compare them for sociodemographic bias with potential for harm. Prior bias measurement datasets are sensitive to perturbations in their manually designed templates, therefore unreliable. To achieve reliability, we introduce the Comprehensive Assessment of Language Model bias (CALM), a benchmark dataset to quantify bias in LMs across three tasks. We integrate 16 existing datasets across different domains, such as Wikipedia and news articles, to filter 224 templates from which we construct a dataset of 78,400 examples. We compare the diversity of CALM with prior datasets on metrics such as average semantic similarity, and variation in template length, and test the sensitivity to small perturbations. We show that our dataset is more diverse and reliable than previous datasets, thus better capture the breadth of linguistic variation required to reliably evaluate model bias. We evaluate 20 large language models including six prominent families of LMs such as Llama-2. In two LM series, OPT and Bloom, we found that larger parameter models are more biased than lower parameter models. We found the T0 series of models to be the least biased. Furthermore, we noticed a tradeoff between gender and racial bias with increasing model size in some model series. The code is available at https://github.com/vipulgupta1011/CALM.

Bayesian predictive inference provides a principled framework for uncertainty quantification, data efficiency, and robust generalization. However, exact inference is often intractable, and scalable approximations may remain computationally expensive or require restrictive modeling assumptions that degrade predictive performance. Prior-Data Fitted and in-context models have recently emerged as an amortized alternative by learning to map datasets directly to predictive distributions, but existing approaches are tightly coupled to the support of the training prior and lack explicit mechanisms for adapting to new priors at test time, resulting in limited robustness under distribution shift. We introduce a multi-task in-context learning framework for amortized hierarchical Bayesian predictive inference that explicitly represents prior information as a prefix of in-context datasets. A transformer trained on sequences of prior and target tasks learns to adapt its predictions across families of priors. On a suite of evaluations with increasing difficulty, including out-of-meta-distribution priors and priors with high-dimensional latent structures, our method matches oracle Bayesian predictors while being orders of magnitude faster. We further demonstrate its practical relevance on a real-world spatiotemporal temperature prediction benchmark. Code is available at https://github.com/martianmartina/multi-task-bayesian-icl/.

Data selection is a key component of efficient instruction tuning for large language models, as recent work has shown that data quality often matters more than data quantity. Accordingly, prior studies have introduced various multi-dimensional heuristics to evaluate and filter instruction data. However, most existing methods rely on static task-agnostic and model-agnostic weighting schemes, which overlook the varying requirements of specific downstream tasks and the differing pre-existing capabilities of models. In this paper, we propose a framework for learning multi-indicator weights that jointly adapts data selection to both the downstream task and the specific model. Our method identifies optimal weight configurations without full-scale fine-tuning by utilizing in-context learning (ICL) signals on compact tiny-validation sets. These signals serve as efficient performance proxies that ensure high-fidelity evaluation at minimal computational cost. Experiments across multiple benchmarks and model families, including Mistral, Qwen, and Llama, show that the approach achieves performance comparable to or exceeding full-dataset tuning while using only 30\% of the training samples on GSM8K. Furthermore, our analysis reveals a trade-off between semantic diversity and logical complexity in reasoning tasks, highlighting the necessity of joint task-model adaptation.

In-Context Learning (ICL) enables Large Language Models (LLMs) to perform tasks by conditioning on input-output examples in the prompt, without requiring any update in model parameters. While widely adopted, it remains unclear whether prompting with multiple examples is the most effective and efficient way to convey task information. In this work, we propose Soft Injection of task embeddings. The task embeddings are constructed only once using few-shot ICL prompts and repeatedly used during inference. Soft injection is performed by softly mixing task embeddings with attention head activations using pre-optimized mixing parameters, referred to as soft head-selection parameters. This method not only allows a desired task to be performed without in-prompt demonstrations but also significantly outperforms existing ICL approaches while reducing memory usage and compute cost at inference time. An extensive evaluation is performed across 57 tasks and 12 LLMs, spanning four model families of sizes from 4B to 70B. Averaged across 57 tasks, our method outperforms 10-shot ICL by 10.2%-14.3% across 12 LLMs. Additional analyses show that our method also serves as an insightful tool for analyzing task-relevant roles of attention heads, revealing that task-relevant head positions selected by our method transfer across similar tasks but not across dissimilar ones -- underscoring the task-specific nature of head functionality. Our soft injection method opens a new paradigm for reducing prompt length and improving task performance by shifting task conditioning from the prompt space to the activation space.

Large language model (LLM) agents perform strongly on short- and mid-horizon tasks, but often break down on long-horizon tasks that require extended, interdependent action sequences. Despite rapid progress in agentic systems, these long-horizon failures remain poorly characterized, hindering principled diagnosis and comparison across domains. To address this gap, we introduce HORIZON, an initial cross-domain diagnostic benchmark for systematically constructing tasks and analyzing long-horizon failure behaviors in LLM-based agents. Using HORIZON, we evaluate state-of-the-art (SOTA) agents from multiple model families (GPT-5 variants and Claude models), collecting 3100+ trajectories across four representative agentic domains to study horizon-dependent degradation patterns. We further propose a trajectory-grounded LLM-as-a-Judge pipeline for scalable and reproducible failure attribution, and validate it with human annotation on trajectories, achieving strong agreement (inter-annotator κ=0.61; human-judge κ=0.84). Our findings offer an initial methodological step toward systematic, cross-domain analysis of long-horizon agent failures and offer practical guidance for building more reliable long-horizon agents. We release our project website at https://xwang2775.github.io/horizon-leaderboard/{HORIZON Leaderboard} and welcome contributions from the community.

Large language models (LLMs) exploit in-context learning (ICL) to solve tasks with only a few demonstrations, but its mechanisms are not yet well-understood. Some works suggest that LLMs only recall already learned concepts from pre-training, while others hint that ICL performs implicit learning over demonstrations. We characterize two ways through which ICL leverages demonstrations. Task recognition (TR) captures the extent to which LLMs can recognize a task through demonstrations -- even without ground-truth labels -- and apply their pre-trained priors, whereas task learning (TL) is the ability to capture new input-label mappings unseen in pre-training. Using a wide range of classification datasets and three LLM families (GPT-3, LLaMA and OPT), we design controlled experiments to disentangle the roles of TR and TL in ICL. We show that (1) models can achieve non-trivial performance with only TR, and TR does not further improve with larger models or more demonstrations; (2) LLMs acquire TL as the model scales, and TL's performance consistently improves with more demonstrations in context. Our findings unravel two different forces behind ICL and we advocate for discriminating them in future ICL research due to their distinct nature.

Media is full of false claims. Even Oxford Dictionaries named "post-truth" as the word of 2016. This makes it more important than ever to build systems that can identify the veracity of a story, and the kind of discourse there is around it. RumourEval is a SemEval shared task that aims to identify and handle rumours and reactions to them, in text. We present an annotation scheme, a large dataset covering multiple topics - each having their own families of claims and replies - and use these to pose two concrete challenges as well as the results achieved by participants on these challenges.

Detecting machine-generated text is essential for transparency and accountability when deploying large language models (LLMs). Among detection approaches, watermarking is a statistically reliable method by design -- it embeds detectable signals into LLM outputs by biasing their token distributions. However, it has been reported that watermarked LLMs often perform worse on downstream tasks. We propose PUPPET, a framework that fine-tunes an LLM via reinforcement learning to generate text that is both more detectable and better performing on downstream tasks. We use two reward functions: a detector that outputs a machine-class likelihood and an evaluator that measures a task-specific metric. Experiments on long-form QA, summarization, and essay writing show that LLMs trained with PUPPET achieve high detectability competitive with watermarking methods while outperforming them on downstream tasks. The analysis shows that this optimization can be performed efficiently with only a few thousand samples in 1--2 GPU hours. Moreover, these gains are consistent across out-of-domain tasks, different LLM families, and model sizes, and are even robust to paraphrasing attacks.

We introduce Tencent WorkBuddy Bench, a multi-domain evaluation suite for coding agents; this report documents its construction methodology, scoring protocol, and a cross-model leaderboard. At its core is a unified evaluation framework for constructing and running distribution-informed coding-agent tasks across four work domains - Code, Web, Office, and Security. Rather than adapting public issue text, every task is reverse-engineered from a real commit, pull request, or business scenario and rewritten as a short, colloquial, role-played request, so that a task's prompt is not recoverable by web-searching the underlying issue, pull request, or commit thread. Because the dataset is released openly - task directories, environment images, evaluation harness, tests, and reference solutions - contamination resistance rests on this construction together with dataset versioning rather than on secrecy. The four subsets - repository-level engineering, front-end development, office and business workflows, and red-/blue-team security - probe complementary facets of real work, each with its own verification style. All are packaged in a uniform task-directory format and run, under a uniform and reproducible protocol, on two agent harnesses (CodeBuddy Code and Claude Code); the full open release makes the benchmark reproducible end to end and directly auditable, since any third party can re-run each task and inspect its content. Because each subset uses a different scoring instrument, scores are not comparable across subsets and the suite reports no suite-wide average. We report a cross-model leaderboard across several model families.

We present our shared task on text-based emotion detection, covering more than 30 languages from seven distinct language families. These languages are predominantly low-resource and spoken across various continents. The data instances are multi-labeled into six emotional classes, with additional datasets in 11 languages annotated for emotion intensity. Participants were asked to predict labels in three tracks: (a) emotion labels in monolingual settings, (b) emotion intensity scores, and (c) emotion labels in cross-lingual settings. The task attracted over 700 participants. We received final submissions from more than 200 teams and 93 system description papers. We report baseline results, as well as findings on the best-performing systems, the most common approaches, and the most effective methods across various tracks and languages. The datasets for this task are publicly available.

Large language models (LLMs) exhibit in-context learning abilities which enable the same model to perform several tasks without any task-specific training. In contrast, traditional adaptation approaches, such as fine-tuning, modify the underlying models for each specific task. In-context learning, however, consistently underperforms task-specific tuning approaches even when presented with the same examples. While most existing approaches (e.g., prompt engineering) focus on the LLM's learned representations to patch this performance gap, our analysis actually reveal that LLM representations contain sufficient information to make good predictions. As such, we focus on the LLM's reasoning abilities and demonstrate that this performance gap exists due to their inability to perform simple probabilistic reasoning tasks. This raises an intriguing question: Are LLMs actually capable of learning how to reason in a task-agnostic manner? We answer this in the affirmative and propose TART which generically improves an LLM's reasoning abilities using a synthetically trained Transformer-based reasoning module. TART trains this reasoning module in a task-agnostic manner using only synthetic logistic regression tasks and composes it with an arbitrary real-world pre-trained model without any additional training. With a single inference module, TART improves performance across different model families (GPT-Neo, Pythia, BLOOM), model sizes (100M - 6B), tasks (14 NLP binary classification tasks), and even across different modalities (audio and vision). Additionally, on the RAFT Benchmark, TART improves GPT-Neo (125M)'s performance such that it outperforms BLOOM (176B), and is within 4% of GPT-3 (175B). Our code and models are available at https://github.com/HazyResearch/TART .

Language models can perceive visual properties of text despite receiving only sequences of tokens-we mechanistically investigate how Claude 3.5 Haiku accomplishes one such task: linebreaking in fixed-width text. We find that character counts are represented on low-dimensional curved manifolds discretized by sparse feature families, analogous to biological place cells. Accurate predictions emerge from a sequence of geometric transformations: token lengths are accumulated into character count manifolds, attention heads twist these manifolds to estimate distance to the line boundary, and the decision to break the line is enabled by arranging estimates orthogonally to create a linear decision boundary. We validate our findings through causal interventions and discover visual illusions--character sequences that hijack the counting mechanism. Our work demonstrates the rich sensory processing of early layers, the intricacy of attention algorithms, and the importance of combining feature-based and geometric views of interpretability.

Real-time safety filtering for large language model (LLM) applications requires classifiers that can detect unsafe prompts, toxic language, jailbreak attempts, and unsafe responses without the cost profile of large guardrail models, and that can distinguish benign sensitive text from genuinely covert harmful content. In this paper, we introduce Opir, a family of encoder-based guardrail models built on the GLiClass architecture. Opir includes multi-task models for binary safe/unsafe classification, multi-label toxicity classification, jailbreak classification, and zero-shot unsafe prompt and response categorization. We also release edge variants with fewer than 100M parameters dedicated to binary safe/unsafe categorization. The models are trained on a three-level taxonomy containing 996 categories across 16 top-level labels, 126 mid-level labels, and 854 leaf labels. Opir's training data combines taxonomy-grounded unsafe prompts, adversarially mined hard negatives, benign safety-preserving examples, generated response examples, multilingual translations, and portions of the Aegis2 and WildGuard training subsets. We also open-sourced an evaluation harness that supports GLiClass and GLiNER2 backends as well as decoder-based models, and covers binary safety classification, multi-label categorization, toxicity, jailbreak detection, prompt safety, response safety, response refusal, and prompt subcategory views across public benchmark families. Across an expanded comparison spanning 12 safety-classification tasks and 17 category tasks against eight contemporary guardrail systems -- including both GLiNER2-based and generative guardrail models -- Opir variants are competitive on or ahead of the strongest open-weight baselines on the majority of benchmark datasets while operating with a substantially smaller deployment footprint.

Retrieval heads, attention heads that copy information from earlier context to the current position, have been proposed as the mechanistic substrate for long-context recall. Rotary position embeddings (RoPE) rotate queries and keys by frequencies decaying with a base hyperparameter theta, and a natural hypothesis is that this rotation either prevents retrieval heads from forming or degrades their function. We test both across four open-weight 7-8B models spanning multi-head and grouped-query attention and a 100x range of theta, using paired-seed needle-in-a-haystack tests, layer-clustered permutation, and causal head-masking. (i) Retrieval heads are causally necessary: masking the 87 detected heads in OLMo-2 collapses recall from 1.00 to 0.00, while masking matched random heads has no effect; this replicates in Qwen. (ii) Higher theta does not reduce retrieval-head count (LLaMA-3.1 at theta=500K has 47 heads vs LLaMA-2 at theta=10K with 42), refuting the prevention hypothesis. (iii) The norm-utility relation is family-specific and significant in opposite directions (Qwen d=-0.49, OLMo d=+0.50, both significant; LLaMA null); since OLMo and LLaMA-3.1 share theta=500K yet differ, the effect is not theta-driven. (iv) Building on Chiang and Yogatama (2025), a controlled patch shows that zeroing the lowest-frequency RoPE dimensions of retrieval heads degrades recall dose-dependently (1.00 to 0.18 when 32 of 128 dimensions are zeroed, vs 0.98 for random dimensions); the effect is head-specific and task-specific. The causal variable is RoPE frequency, not norm-utility. The direction holds in all five models patched (OLMo-2, Qwen2.5-7B/14B, Gemma-2, Mistral) across four lineages and two scales. We do not claim cross-model magnitude. Code and a paired-seed harness are released.

Transferring learned patterns from pretrained neural language models has been shown to significantly improve effectiveness across a variety of language-based tasks, meanwhile further tuning on intermediate tasks has been demonstrated to provide additional performance benefits, provided the intermediate task is sufficiently related to the target task. However, how to identify related tasks is an open problem, and brute-force searching effective task combinations is prohibitively expensive. Hence, the question arises, are we able to improve the effectiveness and efficiency of tasks with no training examples through selective fine-tuning? In this paper, we explore statistical measures that approximate the divergence between domain representations as a means to estimate whether tuning using one task pair will exhibit performance benefits over tuning another. This estimation can then be used to reduce the number of task pairs that need to be tested by eliminating pairs that are unlikely to provide benefits. Through experimentation over 58 tasks and over 6,600 task pair combinations, we demonstrate that statistical measures can distinguish effective task pairs, and the resulting estimates can reduce end-to-end runtime by up to 40%.

Multi-task optimization is a powerful approach for solving a large number of tasks in parallel. However, existing algorithms face distinct limitations: Population-based methods scale poorly and remain underexplored for large task sets. Approaches that do scale beyond a thousand tasks are mostly MAP-Elites variants and rely on a fixed, discretized archive that disregards the topology of the task space. We introduce MONET (Multi-Task Optimization over Networks of Tasks), a multi-task optimization algorithm that models the task space as a graph: tasks are nodes, and edges connect tasks in the task parameter space. This representation enables knowledge transfer between tasks and remains tractable for high-dimensional problems while exploiting the topology of the task space. MONET combines social learning, which generates candidates from neighboring nodes via crossover, with individual learning, which refines a node's own solution independently via mutation. We evaluate MONET on four domains (archery, arm, and cartpole with 5,000 tasks each; hexapod with 2,000 tasks) and show that it matches or exceeds the performance of existing MAP-Elites-based baselines across all four domains.

Task arithmetic, representing downstream tasks through linear operations on task vectors, has emerged as a simple yet powerful paradigm for transferring knowledge across diverse settings. However, maintaining a large collection of task vectors introduces scalability challenges in both storage and computation. We propose Task Vector Bases, a framework compressing T task vectors into M < T basis vectors while preserving the functionality of task arithmetic. By representing each task vector as a structured linear combination of basis atoms, our approach supports standard operations such as addition, negation, as well as more advanced arithmetic ones. The framework is orthogonal to other efficiency-oriented improvements in task arithmetic and can be used in combination with them. We provide theoretical analysis showing that basis compression retains addition generalization guarantees and enables principled unlearning, with error bounds depending on reconstruction quality. Empirically, our proposed basis construction methods consistently outperform heuristic basis construction baselines and, in some cases, even surpass the performance of full task vector collections across diverse downstream applications while reducing storage and computational requirements. The code is available at https://github.com/uiuctml/TaskVectorBasis.

The distribution of subpopulations is an important property hidden within a dataset. Uncovering and analyzing the subpopulation distribution within datasets provides a comprehensive understanding of the datasets, standing as a powerful tool beneficial to various downstream tasks, including Dataset Subpopulation Organization, Subpopulation Shift, and Slice Discovery. Despite its importance, there has been no work that systematically explores the subpopulation distribution of datasets to our knowledge. To address the limitation and solve all the mentioned tasks in a unified way, we introduce a novel concept of subpopulation structures to represent, analyze, and utilize subpopulation distributions within datasets. To characterize the structures in an interpretable manner, we propose the Subpopulation Structure Discovery with Large Language Models (SSD-LLM) framework, which employs world knowledge and instruction-following capabilities of Large Language Models (LLMs) to linguistically analyze informative image captions and summarize the structures. Furthermore, we propose complete workflows to address downstream tasks, named Task-specific Tuning, showcasing the application of the discovered structure to a spectrum of subpopulation-related tasks, including dataset subpopulation organization, subpopulation shift, and slice discovery. Furthermore, we propose complete workflows to address downstream tasks, named Task-specific Tuning, showcasing the application of the discovered structure to a spectrum of subpopulation-related tasks, including dataset subpopulation organization, subpopulation shift, and slice discovery.

Efficiently and accurately translating a corpus into a low-resource language remains a challenge, regardless of the strategies employed, whether manual, automated, or a combination of the two. Many Christian organizations are dedicated to the task of translating the Holy Bible into languages that lack a modern translation. Bible translation (BT) work is currently underway for over 3000 extremely low resource languages. We introduce the eBible corpus: a dataset containing 1009 translations of portions of the Bible with data in 833 different languages across 75 language families. In addition to a BT benchmarking dataset, we introduce model performance benchmarks built on the No Language Left Behind (NLLB) neural machine translation (NMT) models. Finally, we describe several problems specific to the domain of BT and consider how the established data and model benchmarks might be used for future translation efforts. For a BT task trained with NLLB, Austronesian and Trans-New Guinea language families achieve 35.1 and 31.6 BLEU scores respectively, which spurs future innovations for NMT for low-resource languages in Papua New Guinea.

Multimodal large language models (MLLMs) generate responses autoregressively, integrating visual and linguistic information in an evolving context. Prior work on interpretability has focused on individual layers and circuits (where), leaving the token-level dynamics of multimodal computation during generation (when) underexplored. We address this gap and study attention shifts as per semantic role; tracking model attention to image, text, instruction, and previously generated tokens, One Token at a Time (OTaT). We introduce multimodal tasks that require explicit switching between visual and textual context within a single response. Across two mainstream model families and four open-weight MLLMs of varying sizes, we establish consistent patterns: attention to image peaks at tokens requiring image-derived information, instruction tokens are revisited during task transitions, and attention to previously generated tokens increases as the generation progresses. Causal attention blocking interventions validate the functional role of these trends. We profile model behavior under disrupted attention and observe responses falling back to language priors, or exhibiting cross-modal leakage, denial, or recovery. Finally, informed of the attention dynamics through our novel analysis, we propose a simple test-time intervention to boost attention to the relevant modality at the right time, significantly improving multimodal task performance.

In order for large language models to be useful across the globe, they are fine-tuned to follow instructions on multilingual data. Despite the ubiquity of such post-training, a clear understanding of the dynamics that enable cross-lingual transfer remains elusive. This study examines cross-lingual transfer (CLT) dynamics in realistic post-training settings. We study two model families of up to 35B parameters in size trained on carefully controlled mixtures of multilingual data on three generative tasks with varying levels of complexity (summarization, instruction following, and mathematical reasoning) in both single-task and multi-task instruction tuning settings. Overall, we find that the dynamics of cross-lingual transfer and multilingual performance cannot be explained by isolated variables, varying depending on the combination of post-training settings. Finally, we identify the conditions that lead to effective cross-lingual transfer in practice.

Unstructured sparsity is now natively accelerated by recent GPU kernels and dataflow hardware, shifting the bottleneck from inference execution to the pruning algorithm. State-of-the-art methods for unstructured LLM pruning are layer-wise surrogates derived from the Optimal Brain Surgeon principle, and they sacrifice end-to-end accuracy, especially under aggressive sparsity. End-to-end alternatives such as MaskLLM and PATCH show that learnable masks can close this gap, but their categorical-over-patterns parameterization scales with the number of valid masks per row and does not port to the unstructured setting. We introduce LEAP, which replaces this intractable parameterization with a per-weight Bernoulli-via-Gumbel-sigmoid relaxation that makes end-to-end unstructured mask learning tractable. Across five LLM families from 0.5B to 8B parameters at 50% and 60% sparsity, LEAP improves six-task average zero-shot accuracy by +2.59 points on average over ADMM, the best layer-wise baseline in our sweep.

Best-of-N inference scaling (drawing N candidate answers from a language model and returning the one a reward model ranks highest) improves accuracy by an amount that varies across models, but predicting that amount in advance currently requires running the procedure end-to-end. Prior work links cheap statistics of a model's sampled outputs and validation-set correctness (how often samples agree, how diverse they are, how confident the model is, and where correct samples appear) to model behavior, but does not isolate which of these form a stable, compact predictor of best-of-N gain. We fit ridge predictors on features computed from a single labeled validation-set sampling pass, use bootstrap-Lasso as a stability analysis of the candidate feature set, and give a concentration analysis with an explicit linear-approximation residual. Across three base-model families, six post-training methods, and math and reasoning task domains, the stability analysis identifies a strict three-feature core spanning prompt-level agreement spread, label-assisted first-correct-sample position, and completion-length variance; a compact ridge predictor built from this core plus an entropy add-on reaches Spearman ρ= 0.90 with actual best-of-N gain under a reward-model verifier. The intended use is labeled validation-set screening of candidate configurations before paying the full reward-model scoring cost.

Equipping large language models (LLMs) with search engines via reinforcement learning (RL) has emerged as an effective approach for building search agents. However, overreliance on search introduces unnecessary cost and risks exposure to noisy or malicious content, while relying solely on parametric knowledge risks hallucination. The central challenge is to develop agents that adaptively balance parametric knowledge with external search, invoking search only when necessary. Prior work mitigates search overuse by shaping rewards around the number of tool calls. However, these penalties require substantial reward engineering, provide ambiguous credit assignment, and can be exploited by agents that superficially reduce calls. Moreover, evaluating performance solely through call counts conflates necessary and unnecessary search, obscuring the measurement of true adaptive behavior. To address these limitations, we first quantify the self-knowledge awareness of existing search agents via an F1-based decision metric, revealing that methods such as Search-R1 often overlook readily available parametric knowledge. Motivated by these findings, we propose AdaSearch, a simple two-stage, outcome-driven RL framework that disentangles problem solving from the decision of whether to invoke search, and makes this decision process explicit and interpretable. This transparency is crucial for high-stakes domains such as finance and medical question answering, yet is largely neglected by prior approaches. Experiments across multiple model families and sizes demonstrate that AdaSearch substantially improves knowledge-boundary awareness, reduces unnecessary search calls, preserves strong task performance, and offers more transparent, interpretable decision behaviors.

Multimodal Large Language Models (MLLMs) excel in numerous vision-language tasks yet suffer from hallucinations, producing content inconsistent with input visuals, that undermine reliability in precision-sensitive domains. This issue stems from a fundamental problem of visual neglect, where models fail to adequately prioritize input images. Existing methods typically alleviate hallucinations by intervening in the attention score or output logits, focusing on "how to intervene" but overlooking the prerequisite "when to intervene", which leads to the "over-intervention" problem and subsequently introduces new hallucinations and unnecessary computational overhead. To address this gap, we first investigate the mechanism of visual neglect and reveal it can be accurately detected via head-level activation patterns in MLLMs. We thus propose V-ITI, a lightweight visual inference-time intervention framework integrating a Visual Neglect Detector that identifies visual neglect via head-level discriminative probes and a Visual Recall Intervenor that modulates activations with prestored visual activation information only when the visual neglect is detected. Extensive experiments across eight benchmarks and different MLLM families demonstrate that V-ITI consistently mitigates vision-related hallucinations while preserving general task performance.

There has been a surge of interest in assistive wearable agents: agents embodied in wearable form factors (e.g., smart glasses) who take assistive actions toward a user's goal/query (e.g. "Where did I leave my keys?"). In this work, we consider the important complementary problem of inferring that goal from multi-modal contextual observations. Solving this "goal inference" problem holds the promise of eliminating the effort needed to interact with such an agent. This work focuses on creating WAGIBench, a strong benchmark to measure progress in solving this problem using vision-language models (VLMs). Given the limited prior work in this area, we collected a novel dataset comprising 29 hours of multimodal data from 348 participants across 3,477 recordings, featuring ground-truth goals alongside accompanying visual, audio, digital, and longitudinal contextual observations. We validate that human performance exceeds model performance, achieving 93% multiple-choice accuracy compared with 84% for the best-performing VLM. Generative benchmark results that evaluate several families of modern vision-language models show that larger models perform significantly better on the task, yet remain far from practical usefulness, as they produce relevant goals only 55% of the time. Through a modality ablation, we show that models benefit from extra information in relevant modalities with minimal performance degradation from irrelevant modalities.

Large Language Models (LLMs), trained on extensive web-scale corpora, have demonstrated remarkable abilities across diverse tasks, especially as they are scaled up. Nevertheless, even state-of-the-art models struggle in certain cases, sometimes failing at problems solvable by young children, indicating that traditional notions of task complexity are insufficient for explaining LLM capabilities. However, exploring LLM capabilities is complicated by the fact that most widely-used models are also "instruction-tuned" to respond appropriately to prompts. With the goal of disentangling the factors influencing LLM performance, we investigate whether instruction-tuned models possess fundamentally different capabilities from base models that are prompted using in-context examples. Through extensive experiments across various model families, scales and task types, which included instruction tuning 90 different LLMs, we demonstrate that the performance of instruction-tuned models is significantly correlated with the in-context performance of their base counterparts. By clarifying what instruction-tuning contributes, we extend prior research into in-context learning, which suggests that base models use priors from pretraining data to solve tasks. Specifically, we extend this understanding to instruction-tuned models, suggesting that their pretraining data similarly sets a limiting boundary on the tasks they can solve, with the added influence of the instruction-tuning dataset.

Human-object interactions with articulated objects are common in everyday life. Despite much progress in single-view 3D reconstruction, it is still challenging to infer an articulated 3D object model from an RGB video showing a person manipulating the object. We canonicalize the task of articulated 3D human-object interaction reconstruction from RGB video, and carry out a systematic benchmark of five families of methods for this task: 3D plane estimation, 3D cuboid estimation, CAD model fitting, implicit field fitting, and free-form mesh fitting. Our experiments show that all methods struggle to obtain high accuracy results even when provided ground truth information about the observed objects. We identify key factors which make the task challenging and suggest directions for future work on this challenging 3D computer vision task. Short video summary at https://www.youtube.com/watch?v=5tAlKBojZwc

Language models have been shown to exhibit positive scaling, where performance improves as models are scaled up in terms of size, compute, or data. In this work, we introduce NeQA, a dataset consisting of questions with negation in which language models do not exhibit straightforward positive scaling. We show that this task can exhibit inverse scaling, U-shaped scaling, or positive scaling, and the three scaling trends shift in this order as we use more powerful prompting methods or model families. We hypothesize that solving NeQA depends on two subtasks: question answering (task 1) and negation understanding (task 2). We find that task 1 has linear scaling, while task 2 has sigmoid-shaped scaling with an emergent transition point, and composing these two scaling trends yields the final scaling trend of NeQA. Our work reveals and provides a way to analyze the complex scaling trends of language models.

Large language models (LLMs) can now solve complex problems through long chain-of-thought (CoT) reasoning, but the trade-off between performance and token cost remains a central challenge. To address this issue, supervised fine-tuning (SFT) often uses compressed reasoning data, where CoT traces are shortened into compact forms. However, the effect of such compressed reasoning data on post-training remains poorly understood. In this paper, we propose a taxonomy of CoT consisting of Explicit CoT, which outputs all operations without aggregation, Composed CoT, which combines multiple operations into a single step, and Implicit CoT, which omits intermediate operations. We construct a synthetic compositional reasoning task that allows controlled variation of difficulty, compression granularity, and data size, and conducted a comprehensive set of experiments across different model families and sizes. Notably, we find that (i) coarser CoT requires more SFT data, (ii) compared with Explicit CoT, Composed CoT and Implicit CoT benefit more from data scaling, while Composed CoT benefits from data repetition and Implicit CoT tends to lead to memorization, (iii) unlike SFT, subsequent reinforcement learning (RL) with verifiable rewards (RLVR) decomposes compressed steps learned during SFT, and (iv) unidirectional CoT ordering shows stronger generalization on longer sequential tasks. Our findings provide implications for CoT design under data resource constraints and offer important insights into the mechanisms of SFT and RL in LLM post-training.

Multimodal agents repeatedly re-examine the same video frames, UI screenshots, and rendered artifacts as their context window slides and reasoning iterates, yet every look-back re-encodes from scratch, because prefix caches serve reuse only at a fixed leading position. We show this recompute is avoidable, and identify exactly what naive KV reuse loses: the cross-chunk conditioning a chunk absorbs from its neighbours. This loss is asymmetric. The direct readout of a cached chunk is recovered exactly and for free by the standard state-merge. What remains is a diffuse, low-rank residue concentrated in deep layers, invisible to single-hop retrieval but precisely what multi-hop reasoning binds on. Blind reuse therefore leaves single-hop recall intact while halving multi-hop accuracy; this is the failure mode prior position-independent caches, designed for single-context or single-image reuse, do not address. We repair it with a small, training-free low-rank conditioning patch stored alongside each position-free chunk. Reuse reduces to one operator across MLA, GQA, and MHA: exact RoPE re-rotation to any target position, plus the patch that restores cross-chunk binding. This makes three window operations cheap: reorder (one patch serves every ordering of a cached set), sliding-window survival (surviving chunks relocate via rotation only, zero re-encode), and recall (an evicted chunk is rehydrated by its patch, never re-encoded). A rank-m patch recovers full task accuracy on cross-chunk-binding benchmarks, MM-NIAH across two attention families and two-page doc-QA, at a fraction of the KV footprint, and reconstructs re-prefill KV to within bf16 rounding in a production SGLang kernel across six backbones. The conditioning signal is strongest in redundant vision and video streams, making our solution most impactful where multimodal agents spend their recompute budget.

We propose a lightweight and single-pass uncertainty quantification method for detecting hallucinations in Large Language Models. The method uses attention matrices to estimate uncertainty without requiring repeated sampling or external models. Specifically, we measure the Kullback-Leibler divergence between each attention head's distribution and a uniform reference distribution, and use these features in a logistic regression probe. Across multiple datasets, task types, and model families, attention divergence is highly predictive of answer correctness and performs competitively with existing uncertainty estimation methods. We find that this signal is concentrated in middle layers and on factual tokens such as named entities and numbers, suggesting that attention dynamics provides an efficient and interpretable white-box signal of model uncertainty.

Large language models (LLMs) can act as both problem solvers and solution verifiers, where the latter select high-quality answers from a pool of solver-generated candidates. This raises the question of under what conditions verification pays off in solver-verifier systems. Prior work has conducted only limited studies of the factors influencing verification performance, focusing primarily on self-verification and examining neither the relationship between solver and verifier model families nor the effects of reasoning post-training. To rectify this, we present a systematic study across 37 models spanning multiple families, sizes, and base vs. post-trained variants, evaluated on 9 benchmarks covering logical reasoning, structured puzzles, symbolic computation, mathematics, commonsense, factual recall, and domain knowledge. In order to support our analysis, we introduce and empirically validate verifier gain, a metric that predicts the performance improvements from test-time verifier-based rejection sampling. Our experiments find that 1) verification across model families is more effective than either self-verification or verification within the same family, and more generally that the benefits of verification decrease as the solver and verifier become more similar, 2) reasoning post-training weakens self-improvement abilities but strengthens cross-family improvement, and 3) some tasks are inherently more amenable to improvement through verification, particularly mathematical and logical tasks.

Large Language Models (LLMs) have demonstrated remarkable in-context learning (ICL) capabilities. In this study, we explore a surprising phenomenon related to ICL: LLMs can perform multiple, computationally distinct ICL tasks simultaneously, during a single inference call, a capability we term "task superposition". We provide empirical evidence of this phenomenon across various LLM families and scales and show that this phenomenon emerges even if we train the model to in-context learn one task at a time. We offer theoretical explanations that this capability is well within the expressive power of transformers. We also explore how LLMs internally compose task vectors during superposition. Furthermore, we show that larger models can solve more ICL tasks in parallel, and better calibrate their output distribution. Our findings offer insights into the latent capabilities of LLMs, further substantiate the perspective of "LLMs as superposition of simulators", and raise questions about the mechanisms enabling simultaneous task execution.

We present VISTA (VIsual Spec-To-App Benchmark), a benchmark for evaluating the end-to-end web-app generation capabilities of LLM-based agents. Unlike prior code generation benchmarks that focus on algorithmic tasks, VISTA targets realistic UI-centric development, where agents must produce functional, visually coherent applications from underspecified inputs. We define five prompt-information conditions that vary along two axes, visual/structural fidelity and stack constraint: (1) text only with free stack choice, (2) text with reference screenshots under three specified stacks, (3) text with reference screenshots under free stack choice, (4) text with screenshots and pruned Figma structure under a single specified stack, and (5) text with screenshots and pruned Figma structure under free stack choice. To enable robust evaluation, each page in the benchmark is manually annotated with interactive UI components and around three visual anchor points, addressing the well-known limitations of script-based testing tools such as Playwright in open-ended code generation settings. Evaluation combines DOM-grounded reference matching, behavior-specific browser tests, and CLIP-based visual similarity, jointly measuring structural alignment, behavioral completeness, and overall visual fidelity. We use VISTA to assess four agent systems drawn from two model families and two harnesses, finding that visual fidelity and functional correctness are partially decoupled across both input conditions and agents, and that agent editing style varies sharply but is largely orthogonal to task quality. VISTA establishes a rigorous and reproducible foundation for advancing agent-based software engineering research. Code is available at https://github.com/kaboider/VISTA_Bench.

Generating accurate and calibrated confidence estimates is critical for deploying LLMs in high-stakes or user-facing applications, and remains an open challenge. Prior research has often framed confidence as a problem of eliciting a model's "self-knowledge", i.e., the ability of an LLM to judge whether its own answers are correct; this approach implicitly assumes that there is some privileged information about the answer's correctness that is accessible to the model itself. However, our experiments reveal that an LLM attempting to predict the correctness of its own outputs generally performs no better than an unrelated LLM. Moreover, we hypothesize that a key factor in building a "Correctness Model" (CM) is exposure to a target model's historical predictions. We propose multiple methods to inject this historical correctness information, creating a Generalized Correctness Model (GCM). We first show that GCMs can be trained on the correctness data from many LLMs and learn patterns for correctness prediction applicable across datasets and models. We then use CMs as a lens for studying the source of correctness prediction ability and its generalization, systematically controlling their training data and finding that answer phrasing is a strong predictor for correctness. We further explore alternative methods of injecting history without training an LLM, finding that including history as in-context examples can help improve correctness prediction, and post-hoc calibration can provide complementary reductions in calibration error. We evaluate GCMs based on Qwen3-8B across 5 model families and the MMLU and TriviaQA datasets, as well as on a downstream selective prediction task, finding that reliable LLM confidence estimation is a generalizable and model-agnostic skill learned by systematically encoding correctness history rather than a model-specific skill reliant on self-introspection.

Frontier LLM agents are automating many business tasks, but their high inference cost makes large-scale deployment unsustainable. Small language models (SLMs) offer a cheaper alternative, yet they typically fall short when swapped into a harness designed for a frontier LLM. We show that for many routine business tasks, SLM agents can match LLM performance at 90% lower cost, when paired with an adapted harness that can be automatically discovered by a meta agent. The key insight is that much of the task difficulty is shared across instances and can be lifted from the model into the harness via tailored instructions, tools, and orchestration loops. To study this systematically, we create a framework that maps agent failure modes to harness adaptation strategies, and build a harness optimizer that automatically discovers effective adaptations from failure trajectories. Across seven business-oriented agentic tasks and three SLM families, we found optimized harnesses significantly improve performance on 16 of 21 task-SLM pairs, with seven pairs closing the SLM-LLM performance gap and the best SLM agent recovering 89.7% of LLM performance at 4% of the cost. Our analysis further shows that adaptation works best for tasks with more repetitive workflows and for SLMs with sufficient base capabilities. Together, these results suggest that harness adaptation can expand the practical deployment range of SLM agents in routine business tasks.

In large language models (LLMs), each block operates on the residual stream to map input token sequences to output token distributions. However, most of the interpretability literature focuses on internal latent representations, leaving token-space dynamics underexplored. The high dimensionality and categoricity of token distributions hinder their analysis, as standard statistical descriptors are not suitable. We show that the entropy of logit-lens predictions overcomes these issues. In doing so, it provides a per-layer scalar, permutation-invariant metric. We introduce Entropy-Lens to distill the token-space dynamics of the residual stream into a low-dimensional signal. We call this signal the entropy profile. We apply our method to a variety of model sizes and families, showing that (i) entropy profiles uncover token prediction dynamics driven by expansion and pruning strategies; (ii) these dynamics are family-specific and invariant under depth rescaling; (iii) they are characteristic of task type and output format; (iv) these strategies have unequal impact on downstream performance, with the expansion strategy usually being more critical. Ultimately, our findings further enhance our understanding of the residual stream, enabling a granular assessment of how information is processed across model depth.

Machine translation (MT) models are known to suffer from gender bias, especially when translating into languages with extensive gendered morphology. Accordingly, they still fall short in using gender-inclusive language, also representative of non-binary identities. In this paper, we look at gender-inclusive neomorphemes, neologistic elements that avoid binary gender markings as an approach towards fairer MT. In this direction, we explore prompting techniques with large language models (LLMs) to translate from English into Italian using neomorphemes. So far, this area has been under-explored due to its novelty and the lack of publicly available evaluation resources. We fill this gap by releasing Neo-GATE, a resource designed to evaluate gender-inclusive en-it translation with neomorphemes. With Neo-GATE, we assess four LLMs of different families and sizes and different prompt formats, identifying strengths and weaknesses of each on this novel task for MT.

As the capability frontier of autonomous agents continues to expand, they are increasingly able to complete specialized tasks through plug-and-play external skills. Yet current benchmarks mostly test whether models can use provided skills, leaving open whether they can discover skills from experience, repair them after failure, and maintain a coherent library over time. We introduce SkillFlow, a benchmark of 166 tasks across 20 families in which task construction within each family follows a Domain-Agnostic Execution Flow (DAEF) that defines an agent workflow framework, allowing these tasks to share a consistent workflow. Agents are evaluated under an Agentic Lifelong Learning protocol in which they begin without skills, solve tasks sequentially within each family, externalize lessons through trajectory- and rubric-driven skill patches, and carry the updated library forward. Experiments reveal a substantial capability gap. For Claude Opus 4.6, lifelong skill evolution improves task success from 62.65% to 71.08% (+8.43 points). However, high skill usage does not necessarily imply high utility: Kimi K2.5 gains only +0.60 points despite 66.87% skill usage, while Qwen-Coder-Next reaches only a 44.58% task completion rate and still regresses relative to the vanilla setting. SkillFlow contributes a structured testbed for this direction and an in-depth empirical analysis of skill discovery, patching, transfer, and their failure modes under lifelong evaluation.

Planning is a fundamental capability for large language models (LLMs) because such complex tasks require models to coordinate goals, constraints, resources, and long-term consequences into executable and verifiable solutions. Existing planning benchmarks, however, usually treat planning data as fixed collections of instances rather than controllable generation targets. This limits scenario coverage, ties difficulty to surface-level proxies rather than structural sources, and offers limited support for scalable generation, automatic verification, or planning-oriented training. We introduce PlanningBench, a framework for generating scalable, diverse, and verifiable planning data for both evaluation and training. PlanningBench starts from real planning scenarios and abstracts practical workflows into a structured taxonomy of more than 30 task types, subtasks, constraint families, and difficulty factors. Guided by this taxonomy, a constraint-driven synthesis pipeline instantiates self-contained planning problems with adaptive difficulty control, quality filtering, and instance-level verification checklists. This shifts planning data construction from fixed benchmark collection to controllable generation while preserving realistic task grounding. We use PlanningBench to evaluate open-source and closed-source frontier LLMs, and find that current models still struggle to produce complete solutions under coupled constraints. Beyond evaluation, reinforcement learning on verified PlanningBench data improves performance on unseen planning benchmarks and broader instruction-following tasks. Further analysis suggests that determinate or well-specified optimal solutions provide clearer reward signals and more stable training dynamics. Overall, PlanningBench provides a controllable source of planning data for diagnosing and improving generalizable planning abilities in LLMs.

Large vision--language models (VLMs) often use a frozen vision backbone, whose image features are mapped into a large language model through a lightweight connector. While transformer-based encoders are the standard visual backbone, we ask whether state space model (SSM) vision backbones can be a strong alternative. We systematically evaluate SSM vision backbones for VLMs in a controlled setting. Under matched ImageNet-1K initialization, the SSM backbone achieves the strongest overall performance across both VQA and grounding/localization. We further adapt both SSM and ViT-family backbones with detection or segmentation training and find that dense-task tuning generally improves performance across families; after this adaptation, the SSM backbone remains competitive while operating at a substantially smaller model scale. We further observe that (i) higher ImageNet accuracy or larger backbones do not reliably translate into better VLM performance, and (ii) some visual backbones are unstable in localization. Based on these findings, we propose stabilization strategies that improve robustness for both backbone families and highlight SSM backbones as a strong alternative to transformer-based vision encoders in VLMs.

Under standard graphical assumptions, the Markov boundary of a target variable is the smallest set of features that renders every other feature redundant. Once the boundary is observed, the target is conditionally independent of the rest of the table. This is a tempting object for tabular prediction, since it names exactly the columns a model should need. Yet modern regressors are still trained on the full feature set. We ask whether the Markov boundary is genuinely useful for prediction on SCM3K, a 3,450-task synthetic SCM benchmark with feature counts from 40 to 1000 and six SCM families, evaluated with six regressors. The answer is more nuanced than the theory suggests. Restricting a regressor to the oracle boundary often improves prediction substantially, and the improvement grows as the feature space becomes larger and sparser. But the natural pipeline of recovering the boundary with causal discovery and training on the recovered mask does not deliver. Existing estimators exhaust the compute budget before reaching the regime where the boundary helps most, and even where they run they rarely beat the full feature set. We trace this to three causes. Discovery optimizes structural recovery rather than prediction. False negatives and false positives carry sharply asymmetric predictive cost. The exact boundary is only one of many feature sets that beat all features. We then develop what these facts imply for prediction-aligned feature selection and for tabular models that learn to use causal structure.

The Superficial Alignment Hypothesis posits that almost all of a language model's abilities and knowledge are learned during pre-training, while post-training is about giving a model the right style and format. We re-examine these claims by empirically studying the scaling behavior of post-training with increasing finetuning examples and evaluating them using objective task-specific standardized benchmarks. Through experiments with the Llama-3, Mistral, and Llama-2 model families of multiple sizes, we observe that, similar to the pre-training scaling laws, post-training task performance scales as a power law against the number of finetuning examples. This power law relationship holds across a broad array of capabilities, including mathematical reasoning, coding, instruction following, and multihop-reasoning. In addition, for tasks like math and multihop reasoning, we observe that a handful of examples merely align the model stylistically but do not saturate performance on the benchmarks. Model performance is instead correlated with its reasoning ability and it improves significantly with more examples, illustrating the need for holistic evaluation programs leveraging objective benchmarks in addition to measurement of alignment to human preferences. We also observe that language models are not necessarily limited to using knowledge learned during pre-training. With appropriate post-training, a model's ability to integrate new knowledge greatly improves on downstream tasks like multihop question-answering. Taken together, these results shed new light on the Superficial Alignment Hypothesis, suggesting that it is, at best, an over-simplification.

Artificial Intelligence (AI) has made incredible progress recently. On the one hand, advanced foundation models like ChatGPT can offer powerful conversation, in-context learning and code generation abilities on a broad range of open-domain tasks. They can also generate high-level solution outlines for domain-specific tasks based on the common sense knowledge they have acquired. However, they still face difficulties with some specialized tasks because they lack enough domain-specific data during pre-training or they often have errors in their neural network computations on those tasks that need accurate executions. On the other hand, there are also many existing models and systems (symbolic-based or neural-based) that can do some domain-specific tasks very well. However, due to the different implementation or working mechanisms, they are not easily accessible or compatible with foundation models. Therefore, there is a clear and pressing need for a mechanism that can leverage foundation models to propose task solution outlines and then automatically match some of the sub-tasks in the outlines to the off-the-shelf models and systems with special functionalities to complete them. Inspired by this, we introduce TaskMatrix.AI as a new AI ecosystem that connects foundation models with millions of APIs for task completion. Unlike most previous work that aimed to improve a single AI model, TaskMatrix.AI focuses more on using existing foundation models (as a brain-like central system) and APIs of other AI models and systems (as sub-task solvers) to achieve diversified tasks in both digital and physical domains. As a position paper, we will present our vision of how to build such an ecosystem, explain each key component, and use study cases to illustrate both the feasibility of this vision and the main challenges we need to address next.

Crossmodal correspondences between sound and taste are well established in psychology and neuroscience, but largely absent from content-based multimedia retrieval. We formalise taste-from-audio prediction as a content-based music information retrieval benchmark over a perceptually validated multi-source corpus, comparing ten frozen audio encoders from the four HEAR families under a shared multi-task regression head, with gated late-fusion as a configurable variant. In order to assess the effectiveness of the models, we compute absolute error and rank correlation. The strongest systems predict the five tastes within a macro RMSE of 0.134; on held-out real music their error is less than half a single rater's deviation from the consensus (RMSE 0.13 vs. 0.28), so the model tracks the group consensus more closely than an average human rater, and well below the previous state of the art baseline (0.219). On absolute error the encoders are statistically flat, with a single VGGish matching the best fusion, but gated late-fusion's advantage is confined to rank correlation (macro Pearson r 0.724 vs. 0.666). Operationalised as a content-based retrieval index, the predicted taste space ranks a 309-item pool far more faithfully than a CLAP-text baseline, which sits at chance; ridge probes and an audio-bandstop knockout read the strongest representations against documented sound-taste correspondences.

Large language models are highly sensitive to prompts, but this sensitivity is usually studied through task-relevant instructions, demonstrations, or reasoning cues. In this paper, we study a different form of prompt sensitivity: whether prompts that are semantically unrelated to the task can nevertheless steer model behavior. We call them spurious prompts and show their surprising efficacy. We also propose a simple black-box search procedure for discovering them. Across reasoning and question-answering benchmarks, using models ranging from 0.8B to 27B parameters and spanning three model families, we show that spurious prompts can improve performance, often matching or outperforming standard prompting baselines and task-aware prompt optimization. We further show that they can steer models toward unintended behaviors, such as repeatedly selecting the first answer option, producing incorrect answers, returning an even, prime or small number without explicitly instructing the model to do so. These findings reveal a new kind of prompt sensitivity: LLMs can be systematically steered by prompts that are unrelated to the task they are asked to solve. Our code is available at https://github.com/Batorskq/spurious

Today's evolving labor markets rely increasingly on recommender systems for hiring, talent management, and workforce analytics, with natural language processing (NLP) capabilities at the core. Yet, research in this area remains highly fragmented. Studies employ divergent ontologies (ESCO, O*NET, national taxonomies), heterogeneous task formulations, and diverse model families, making cross-study comparison and reproducibility exceedingly difficult. General-purpose benchmarks lack coverage of work-specific tasks, and the inherent sensitivity of employment data further limits open evaluation. We present WorkRB (Work Research Benchmark), the first open-source, community-driven benchmark tailored to work-domain AI. WorkRB organizes 13 diverse tasks from 7 task groups as unified recommendation and NLP tasks, including job/skill recommendation, candidate recommendation, similar item recommendation, and skill extraction and normalization. WorkRB enables both monolingual and cross-lingual evaluation settings through dynamic loading of multilingual ontologies. Developed within a multi-stakeholder ecosystem of academia, industry, and public institutions, WorkRB has a modular design for seamless contributions and enables integration of proprietary tasks without disclosing sensitive data. WorkRB is available under the Apache 2.0 license at https://github.com/techwolf-ai/WorkRB.

As multimodal large language models (MLLMs) advance in handling interleaved image-text data, assessing their few-shot learning capabilities remains an open challenge. In this paper, we introduce FewMMBench, a comprehensive benchmark designed to evaluate MLLMs under few-shot conditions, with a focus on In-Context Learning (ICL) and Chain-of-Thought (CoT) prompting. Covering a diverse suite of multimodal understanding tasks, from attribute recognition to temporal reasoning, FewMMBench enables systematic analysis across task types, model families, and prompting strategies. We evaluate 26 open-weight MLLMs from six model families across zero-shot, few-shot, and CoT-augmented few-shot settings. Our findings reveal that instruction-tuned models exhibit strong zero-shot performance but benefit minimally, or even regress, with additional demonstrations or CoT reasoning. Retrieval-based demonstrations and increased context size also yield limited gains. These results highlight FewMMBench as a rigorous testbed for diagnosing and advancing few-shot capabilities in multimodal LLMs. The data is available at: https://huggingface.co/datasets/mustafaa/FewMMBench

As the era of autonomous agents making decisions on behalf of users unfolds, ensuring contextual integrity (CI) -- what is the appropriate information to share while carrying out a certain task -- becomes a central question to the field. We posit that CI demands a form of reasoning where the agent needs to reason about the context in which it is operating. To test this, we first prompt LLMs to reason explicitly about CI when deciding what information to disclose. We then extend this approach by developing a reinforcement learning (RL) framework that further instills in models the reasoning necessary to achieve CI. Using a synthetic, automatically created, dataset of only sim700 examples but with diverse contexts and information disclosure norms, we show that our method substantially reduces inappropriate information disclosure while maintaining task performance across multiple model sizes and families. Importantly, improvements transfer from this synthetic dataset to established CI benchmarks such as PrivacyLens that has human annotations and evaluates privacy leakage of AI assistants in actions and tool calls.

We introduce KFinEval-Pilot, a benchmark suite specifically designed to evaluate large language models (LLMs) in the Korean financial domain. Addressing the limitations of existing English-centric benchmarks, KFinEval-Pilot comprises over 1,000 curated questions across three critical areas: financial knowledge, legal reasoning, and financial toxicity. The benchmark is constructed through a semi-automated pipeline that combines GPT-4-generated prompts with expert validation to ensure domain relevance and factual accuracy. We evaluate a range of representative LLMs and observe notable performance differences across models, with trade-offs between task accuracy and output safety across different model families. These results highlight persistent challenges in applying LLMs to high-stakes financial applications, particularly in reasoning and safety. Grounded in real-world financial use cases and aligned with the Korean regulatory and linguistic context, KFinEval-Pilot serves as an early diagnostic tool for developing safer and more reliable financial AI systems.

In the evolving landscape of natural language processing (NLP), fine-tuning pre-trained Large Language Models (LLMs) with first-order (FO) optimizers like SGD and Adam has become standard. Yet, as LLMs grow {in size}, the substantial memory overhead from back-propagation (BP) for FO gradient computation presents a significant challenge. Addressing this issue is crucial, especially for applications like on-device training where memory efficiency is paramount. This paper proposes a shift towards BP-free, zeroth-order (ZO) optimization as a solution for reducing memory costs during LLM fine-tuning, building on the initial concept introduced by MeZO. Unlike traditional ZO-SGD methods, our work expands the exploration to a wider array of ZO optimization techniques, through a comprehensive, first-of-its-kind benchmarking study across five LLM families (Roberta, OPT, LLaMA, Vicuna, Mistral), three task complexities, and five fine-tuning schemes. Our study unveils previously overlooked optimization principles, highlighting the importance of task alignment, the role of the forward gradient method, and the balance between algorithm complexity and fine-tuning performance. We further introduce novel enhancements to ZO optimization, including block-wise descent, hybrid training, and gradient sparsity. Our study offers a promising direction for achieving further memory-efficient LLM fine-tuning. Codes to reproduce all our experiments are at https://github.com/ZO-Bench/ZO-LLM .

With the advancement of large language models (LLMs), solving complex reasoning tasks has gained increasing attention. Inference-time computation methods (e.g., Best-of-N, beam search, et al.) are particularly valuable as they can enhance reasoning performance without modifying model parameters or requiring additional training. However, these techniques come with implementation challenges, and most existing methods remain at the proof-of-concept stage with limited practical adoption due to their computational complexity and varying effectiveness across different tasks. In this paper, we investigate and benchmark diverse inference-time computation strategies across reasoning tasks of varying complexity. Since most current methods rely on a proposer-verifier pipeline that first generates candidate solutions (e.g., reasoning solutions) and then selects the best one based on reward signals (e.g., RLHF rewards, process rewards), our research focuses on optimizing both candidate solution generation (e.g., instructing prompts, hyperparameters such as temperature and top-p) and reward mechanisms (e.g., self-evaluation, reward types). Through extensive experiments (more than 20,000 A100-80G GPU hours with over 1,000 experiments) across a variety of models (e.g., Llama, Qwen, and Mistral families) of various sizes, our ablation studies reveal that previously overlooked strategies can significantly enhance performance (e.g., tuning temperature can improve reasoning task performance by up to 5%). Furthermore, we establish a standardized benchmark for inference-time computation by systematically evaluating six representative methods across eight reasoning tasks. These findings provide a stronger foundation for future research. The code is available at https://github.com/usail-hkust/benchmark_inference_time_computation_LLM

Recent growth in reinforcement learning (RL) has surfaced a need for diverse, specialized training environments. Hand-curated environments with fixed task and reward difficulties become ineffective signals as model performance improves, and sparse rewards over long horizons induce mode collapse on specific workflows or tool structures. World models that simulate environment states have matched pure rollout performance, making them promising for scaling diversity on-demand. However, autoregressive (AR) world models suffer from a left-to-right bias preventing conditioning on globally interdependent state anchors such as tool schemas, prior turns, and expected outcomes. We (i) formalize text-based world modeling as a steerable transition-dynamics problem decomposed into initial state, task context, tool schemas, domain rules, and steering directives, and (ii) curate 239,403 grounded state-action trajectories spanning nine open-source environments and twelve frontier model families. We compare AR LMs and masked diffusion language models (MDLMs), showing MDLMs, via bidirectional anchor-aware denoising, achieve better coherence, groundedness, and empirically validated rollout diversity than LLMs over 4x their parameter size, at comparable inference latency. We introduce a plug-and-play GRPO training framework with deterministic state checks, and perform zero-shot transfer ablations on three OOD environments (ScienceWorld, ALFWorld, AppWorld) across three 1.2B-7B agent backbones (LFM2.5, Qwen3, Mistral), achieving up to 47% absolute gains over baselines without environment-specific fine-tuning. We further conduct behavioral analysis of failure modes under adversarial scenarios and human evaluation on realism, outcome correctness, and training utility. We open-source our work to encourage research in this direction.

This work explores the problem of generating task graphs of real-world activities. Different from prior formulations, we consider a setting where text transcripts of instructional videos performing a real-world activity (e.g., making coffee) are provided and the goal is to identify the key steps relevant to the task as well as the dependency relationship between these key steps. We propose a novel task graph generation approach that combines the reasoning capabilities of instruction-tuned language models along with clustering and ranking components to generate accurate task graphs in a completely unsupervised manner. We show that the proposed approach generates more accurate task graphs compared to a supervised learning approach on tasks from the ProceL and CrossTask datasets.

Experts in various fields routinely perform methodical writing tasks to plan, organize, and report their work. From a clinician writing a differential diagnosis for a patient, to a teacher writing a lesson plan for students, these tasks are pervasive, requiring to methodically generate structured long-form output for a given input. We develop a typology of methodical tasks structured in the form of a task objective, procedure, input, and output, and introduce DoLoMiTes, a novel benchmark with specifications for 519 such tasks elicited from hundreds of experts from across 25 fields. Our benchmark further contains specific instantiations of methodical tasks with concrete input and output examples (1,857 in total) which we obtain by collecting expert revisions of up to 10 model-generated examples of each task. We use these examples to evaluate contemporary language models highlighting that automating methodical tasks is a challenging long-form generation problem, as it requires performing complex inferences, while drawing upon the given context as well as domain knowledge.

Prompt routing dynamically selects the most appropriate large language model from a pool of candidates for each query, optimizing performance while managing costs. As model pools scale to include dozens of frontier models with narrow performance gaps, existing approaches face significant challenges: manually defined task taxonomies cannot capture fine-grained capability distinctions, while monolithic routers struggle to differentiate subtle differences across diverse tasks. We propose a two-stage routing architecture that addresses these limitations through automated fine-grained task discovery and task-aware quality estimation. Our first stage employs graph-based clustering to discover latent task types and trains a classifier to assign prompts to discovered tasks. The second stage uses a mixture-of-experts architecture with task-specific prediction heads for specialized quality estimates. At inference, we aggregate predictions from both stages to balance task-level stability with prompt-specific adaptability. Evaluated on 10 benchmarks with 11 frontier models, our method consistently outperforms existing baselines and surpasses the strongest individual model while incurring less than half its cost.

We introduce PhysicianBench, a benchmark for evaluating LLM agents on physician tasks grounded in real clinical setting within electronic health record (EHR) environments. Existing medical agent benchmarks primarily focus on static knowledge recall, single-step atomic actions, or action intent without verifiable execution against the environment. As a result, they fail to capture the long-horizon, composite workflows that characterize real clinical systems. PhysicianBench comprises 100 long-horizon tasks adapted from real consultation cases between primary care and subspecialty physicians, with each task independently reviewed by a separate panel of physicians. Tasks are instantiated in an EHR environment with real patient records and accessed through the same standard APIs used by commercial EHR vendors. Tasks span 21 specialties (e.g., cardiology, endocrinology, oncology, psychiatry) and diverse workflow types (e.g., diagnosis interpretation, medication prescribing, treatment planning), requiring an average of 27 tool calls per task. Solving each task requires retrieving data across encounters, reasoning over heterogeneous clinical information, executing consequential clinical actions, and producing clinical documentation. Each task is decomposed into structured checkpoints (670 in total across the benchmark) capturing distinct stages of completion graded by task-specific scripts with execution-grounded verification. Across 13 proprietary and open-source LLM agents, the best-performing model achieves only 46% success rate (pass@1), while open-source models reach at most 19%, revealing a substantial gap between current agent capabilities and the demands of real-world clinical workflows. PhysicianBench provides a realistic and execution-grounded benchmark for measuring progress toward autonomous clinical agents.

Geometric Attention (GA) specifies an attention layer by four independent inputs: a finite carrier (what indices are addressable), an evidence-kernel rule (how masked proto-scores and a link induce nonnegative weights), a probe family (which observables are treated as admissible), and an anchor/update rule (which representative kernel is selected and how it is applied). Probe families induce an operational equivalence relation on kernels and therefore a gauge; anchors select representatives relative to that probe. Under a scalar relational-work representation and a multiplicative compositionality law for evidence, the admissible link family is exponential, yielding Gibbs weights; with row anchoring this includes the softmax kernel family as a subregime. After quotienting unary row/column score fields, the remaining interaction component admits a canonical rank-r normal form (Eckart-Young/SVD); dot-product score charts implement the corresponding low-rank interaction regime. Fixing the carrier and extensionalizing the update yields the standard fixed-token Transformer attention operator; allowing carrier updates yields adaptive-carrier and staged-depth regimes. The operator language also supports multihead/mixed kernels, plan-based anchors (e.g., entropic OT/Sinkhorn), and unary operators (e.g., FFN-style fields) as explicit regime choices. This separates invariant structure from modeling choice, enabling principled comparison and extension of attention mechanisms, and attention-based architectures.

TaskGen is an open-sourced agentic framework which uses an Agent to solve an arbitrary task by breaking them down into subtasks. Each subtask is mapped to an Equipped Function or another Agent to execute. In order to reduce verbosity (and hence token usage), TaskGen uses StrictJSON that ensures JSON output from the Large Language Model (LLM), along with additional features such as type checking and iterative error correction. Key to the philosophy of TaskGen is the management of information/memory on a need-to-know basis. We empirically evaluate TaskGen on various environments such as 40x40 dynamic maze navigation with changing obstacle locations (100% solve rate), TextWorld escape room solving with dense rewards and detailed goals (96% solve rate), web browsing (69% of actions successful), solving the MATH dataset (71% solve rate over 100 Level-5 problems), Retrieval Augmented Generation on NaturalQuestions dataset (F1 score of 47.03%)

As agent capabilities advance, existing benchmarks, such as τ^2-Bench, are becoming increasingly saturated. Yet constructing new benchmark tasks remains complex, costly, and labor-intensive. Moreover, the standard approach, in which scenarios are first written in natural language and then mapped to tool sequences, captures only a narrow subset of the tool-use patterns agents exercise. In this paper, we address these problems by reversing the task construction process. We propose TASTE: Task Synthesis from Tool Sequence Evolution, an automatic method that generates challenging tasks with broader tool-use coverage. TASTE utilizes an Adaptive Contrastive n-gram model trained on LLM-judged validity signals. This enables sampling valid tool sequences that cover a vast range of tool combinations. TASTE then selects representative sequences from the pool via clustering, instantiates them into complete benchmark tasks, and refines them through iterative difficulty evolution. Using TASTE, we construct τ^c-Bench, a challenging extension of the three domains of τ^2-Bench. We evaluate 11 agent/user LLM pairs and find that models nearly saturating τ^2-Bench suffer severe performance drops on our tasks (e.g., Gemini-3-Flash falls from 0.82!-!0.94 to 0.28!-!0.61). Beyond increasing difficulty, our generated tasks more than double the number of unique tool combinations agents must execute. Our results suggest high scores on existing benchmarks often reflect saturation rather than robust task-solving ability. By automating the generation of difficult, high-coverage benchmarks, TASTE enables continuous, scalable evaluation of future agents.

We study the problem of retrieval with instructions, where users of a retrieval system explicitly describe their intent along with their queries. We aim to develop a general-purpose task-aware retrieval system using multi-task instruction tuning, which can follow human-written instructions to find the best documents for a given query. We introduce the first large-scale collection of approximately 40 retrieval datasets with instructions, BERRI, and present TART, a multi-task retrieval system trained on BERRI with instructions. TART shows strong capabilities to adapt to a new retrieval task via instructions and advances the state of the art on two zero-shot retrieval benchmarks, BEIR and LOTTE, outperforming models up to three times larger. We further introduce a new evaluation setup, X^2-Retrieval to better reflect real-world scenarios, where diverse domains and tasks are pooled and a system needs to find documents aligning users' intents. In this setup, TART significantly outperforms competitive baselines, further demonstrating the effectiveness of guiding retrieval with instructions.

Learning discriminative task-specific features simultaneously for multiple distinct tasks is a fundamental problem in multi-task learning. Recent state-of-the-art models consider directly decoding task-specific features from one shared task-generic feature (e.g., feature from a backbone layer), and utilize carefully designed decoders to produce multi-task features. However, as the input feature is fully shared and each task decoder also shares decoding parameters for different input samples, it leads to a static feature decoding process, producing less discriminative task-specific representations. To tackle this limitation, we propose TaskExpert, a novel multi-task mixture-of-experts model that enables learning multiple representative task-generic feature spaces and decoding task-specific features in a dynamic manner. Specifically, TaskExpert introduces a set of expert networks to decompose the backbone feature into several representative task-generic features. Then, the task-specific features are decoded by using dynamic task-specific gating networks operating on the decomposed task-generic features. Furthermore, to establish long-range modeling of the task-specific representations from different layers of TaskExpert, we design a multi-task feature memory that updates at each layer and acts as an additional feature expert for dynamic task-specific feature decoding. Extensive experiments demonstrate that our TaskExpert clearly outperforms previous best-performing methods on all 9 metrics of two competitive multi-task learning benchmarks for visual scene understanding (i.e., PASCAL-Context and NYUD-v2). Codes and models will be made publicly available at https://github.com/prismformore/Multi-Task-Transformer

Multi-Task Learning (MTL) is a powerful technique that has gained popularity due to its performance improvement over traditional Single-Task Learning (STL). However, MTL is often challenging because there is an exponential number of possible task groupings, which can make it difficult to choose the best one, and some groupings might produce performance degradation due to negative interference between tasks. Furthermore, existing solutions are severely suffering from scalability issues, limiting any practical application. In our paper, we propose a new data-driven method that addresses these challenges and provides a scalable and modular solution for classification task grouping based on hand-crafted features, specifically Data Maps, which capture the training behavior for each classification task during the MTL training. We experiment with the method demonstrating its effectiveness, even on an unprecedented number of tasks (up to 100).

We introduce Dynatask: an open source system for setting up custom NLP tasks that aims to greatly lower the technical knowledge and effort required for hosting and evaluating state-of-the-art NLP models, as well as for conducting model in the loop data collection with crowdworkers. Dynatask is integrated with Dynabench, a research platform for rethinking benchmarking in AI that facilitates human and model in the loop data collection and evaluation. To create a task, users only need to write a short task configuration file from which the relevant web interfaces and model hosting infrastructure are automatically generated. The system is available at https://dynabench.org/ and the full library can be found at https://github.com/facebookresearch/dynabench.

Reward machines (RMs) are a recent formalism for representing the reward function of a reinforcement learning task through a finite-state machine whose edges encode subgoals of the task using high-level events. The structure of RMs enables the decomposition of a task into simpler and independently solvable subtasks that help tackle long-horizon and/or sparse reward tasks. We propose a formalism for further abstracting the subtask structure by endowing an RM with the ability to call other RMs, thus composing a hierarchy of RMs (HRM). We exploit HRMs by treating each call to an RM as an independently solvable subtask using the options framework, and describe a curriculum-based method to learn HRMs from traces observed by the agent. Our experiments reveal that exploiting a handcrafted HRM leads to faster convergence than with a flat HRM, and that learning an HRM is feasible in cases where its equivalent flat representation is not.