ccui46/tale_suite_trajectories

TALES Agent Trajectories

Trajectories from TALES: Text Adventure Learning Environment Suite (NeurIPS 2026, Track on Evaluations and Datasets), by Christopher Zhang Cui, Xingdi Yuan, Ziang Xiao, Prithviraj Ammanabrolu and Marc-Alexandre Côté (University of California, San Diego; Microsoft Research Montréal; Johns Hopkins University). Code: https://github.com/microsoft/tale-suite

Agent trajectory data from a text-adventure benchmark suite.

Leaderboard

Top agents ranked by average best normalized score per game across 122 games, each repeated over 5 seeds (610 total). Scores reflect the highest normalized score achieved at any point during each playthrough, not the final score.

Note: The leaderboard covers the 76 configurations evaluated on all 122 games × 5 seeds. The released files also include claude-fable-5 (high, with action tool; one seed per game) and 500-step Jericho-only runs of gpt-6-astra and claude-opus-5.5 (max_steps = 500 in raw/), which are not ranked here. At 500 steps, their Jericho normalized high scores are 0.669 (gpt-6-astra) and 0.570 (claude-opus-5.5). Both 500-step runs use maximum reasoning effort and the full interaction history in context, block save, load, restore and undo commands, and do not state the step budget in the prompt.

Action tool (+tool): Claude models released after claude-opus-4.6, when run without an action tool, often do not stop after their first command and instead continue with imagined later observations, which they then treat as fact. These models are therefore evaluated with an action tool: the system prompt asks the model to submit each command by calling a submit_command tool, and the tool's command argument is sent to the game as the action. These configurations are marked +tool in the leaderboard and use the variant react-high-tool in transcript_id. claude-opus-4.6 was also re-run with the tool as a control; its scores change minimally. claude-opus-5.5 and gpt-6-astra do not show this behavior. claude-opus-5.5 plays with bare-text actions; the 500-step gpt-6-astra runs submit actions through a structured action tool (variant react-max-tool) and were not re-run without it, since the tool changes scores minimally.

Interpreter commands (claude-opus-5.5): in its 500-step runs, claude-opus-5.5 repeatedly and deliberately tried to manipulate the underlying Frotz interpreter, sending dfrotz backslash meta-commands such as \_ (which dfrotz reads as Return) or \help in place of game commands, mostly to get past in-game menus and key-press prompts. The harness for these runs rejects interpreter meta-commands, so 121 of its actions, in 36 of its 270 runs, were blocked. Each blocked step has the action talesunsafeaction and the reply "The rejected action was discarded; this step was wasted.", and still uses one of the 500 steps. The commands it submitted are listed in extras/rejected_actions.jsonl (see Rejected Actions below).

Missing transcript: one transcript, gpt-oss-20b (high), ScienceWorldChemistryMix, seed 202411065, is missing from this release and will be added in a future update. Its score is included in the leaderboard.

Rank	Model	TW	TWX	ALFWorld	SciWorld	Jericho	benchmark
1	claude-opus-5 (high) +tool	1.000	0.968	1.000	0.950	0.222	0.639
2	claude-opus-4.7 (high) +tool	1.000	0.968	1.000	0.963	0.190	0.628
3	gpt-5.4 (high)	1.000	0.947	1.000	0.952	0.169	0.613
4	claude-opus-4.6 (high) +tool	1.000	0.985	1.000	0.909	0.178	0.612
5	claude-opus-4.8 (high) +tool	1.000	0.974	1.000	0.915	0.173	0.610
6	claude-opus-4.6 (high)	1.000	0.957	1.000	0.919	0.162	0.604
7	claude-opus-4.5 (high)	1.000	0.891	1.000	0.940	0.168	0.603
8	o3 (medium)	1.000	0.919	0.883	0.930	0.157	0.588
9	gpt-5.6-terra (high)	0.987	0.819	1.000	0.946	0.146	0.584
10	gpt-5.1 (high)	0.995	0.898	0.917	0.902	0.161	0.583
11	o3 (high)	1.000	0.896	0.817	0.931	0.161	0.580
12	claude-sonnet-4.6 (high)	1.000	0.905	1.000	0.887	0.130	0.575
13	gpt-5 (high)	1.000	0.755	0.933	0.918	0.172	0.575
14	kimi-k2.6 (high)	1.000	0.885	0.967	0.869	0.144	0.571
15	gpt-5.6-luna (high)	1.000	0.871	0.883	0.914	0.099	0.552
16	o3 (low)	0.991	0.898	0.700	0.883	0.142	0.548
17	claude-4-sonnet	0.996	0.784	0.917	0.870	0.123	0.543
18	kimi-k2.5 (high)	0.991	0.830	0.883	0.864	0.112	0.539
19	claude-3.7-sonnet (1024)	0.973	0.913	0.833	0.765	0.125	0.525
20	claude-3.7-sonnet	0.973	0.958	0.817	0.724	0.130	0.521
21	claude-3.5-sonnet-latest	0.955	0.816	0.750	0.823	0.096	0.504
22	gpt-4.1	0.953	0.925	0.833	0.761	0.068	0.499
23	grok-4.3 (high)	0.964	0.611	0.967	0.832	0.090	0.498
24	Qwen3.5-122B-A10B (1024)	0.914	0.834	0.633	0.765	0.076	0.468
25	Qwen3.5-27B (1024)	0.920	0.861	0.750	0.685	0.077	0.465
26	gpt-5-mini (high)	0.947	0.619	0.617	0.827	0.095	0.465
27	Qwen3.6-27B (1024)	0.940	0.895	0.700	0.685	0.062	0.459
28	o1 (high)	0.978	0.702	0.283	0.801	0.103	0.443
29	claude-sonnet-4.5	1.000	0.714	0.833	0.567	0.086	0.435
30	gpt-4o	0.836	0.806	0.567	0.614	0.056	0.406
31	claude-3.5-haiku	0.949	0.797	0.267	0.673	0.050	0.396
32	Qwen3.6-35B-A3B (1024)	0.817	0.734	0.417	0.598	0.028	0.364
33	Llama-3.1-405B-Instruct	0.909	0.792	0.317	0.518	0.061	0.364
34	gemma-4-31b-it (1024)	0.864	0.576	0.600	0.467	0.069	0.351
35	gemini-2.0-flash	0.808	0.761	0.200	0.571	0.054	0.350
36	Qwen3-32B (1024)	0.795	0.689	0.483	0.498	0.040	0.343
37	Llama-3.3-70B-Instruct	0.696	0.772	0.150	0.551	0.045	0.328
38	Llama-3.1-70B-Instruct	0.656	0.819	0.083	0.519	0.053	0.320
39	Qwen2.5-72B-Instruct	0.765	0.838	0.367	0.350	0.029	0.307
40	Qwen3.5-9B (1024)	0.733	0.726	0.183	0.474	0.035	0.305
41	Mistral-Large-Instruct-2407	0.824	0.683	0.067	0.461	0.058	0.303
42	o3-mini (high)	0.839	0.613	0.117	0.481	0.050	0.301
43	MiniMax-M2.7 (1024)	0.640	0.304	0.217	0.553	0.064	0.278
44	gpt-4.1-mini	0.621	0.745	0.050	0.419	0.034	0.271
45	Llama-3.3-Nemotron-Super-49B-v1 (1024)	0.658	0.789	0.167	0.265	0.028	0.251
46	gpt-5.4-nano (high)	0.697	0.679	0.050	0.358	0.022	0.249
47	Qwen3-30B-A3B (1024)	0.555	0.582	0.167	0.315	0.030	0.229
48	gpt-4o-mini	0.566	0.736	0.000	0.272	0.018	0.218
49	claude-haiku-4.5	0.355	0.542	0.000	0.305	0.054	0.199
50	Llama-4-Scout-17B-16E-Instruct	0.411	0.684	0.000	0.270	0.019	0.198
51	Qwen3.5-4B (1024)	0.377	0.638	0.000	0.279	0.018	0.191
52	gpt-5-nano (high)	0.500	0.413	0.017	0.320	0.017	0.183
53	gpt-oss-20b (high)	0.606	0.451	0.000	0.257	0.016	0.179
54	Llama-4-Maverick-17B-128E-Instruct-FP8	0.435	0.561	0.083	0.114	0.020	0.155
55	Mistral-Small-Instruct-2409	0.561	0.273	0.000	0.244	0.014	0.148
56	DeepSeek-V4-Flash (1024)	0.313	0.427	0.017	0.177	0.046	0.147
57	Llama-3.1-8B-Instruct	0.297	0.503	0.000	0.158	0.023	0.139
58	DeepSeek-R1 (1024)	0.371	0.386	0.000	0.158	0.010	0.124
59	Qwen2.5-7B-Instruct	0.276	0.456	0.000	0.126	0.007	0.117
60	Llama-3.2-3B-Instruct	0.214	0.420	0.000	0.100	0.015	0.104
61	phi-4	0.209	0.438	0.000	0.089	0.015	0.103
62	gpt-4.1-nano	0.128	0.387	0.000	0.094	0.036	0.100
63	Mistral-Small-24B-Instruct-2501	0.158	0.230	0.000	0.158	0.014	0.088
64	DeepSeek-R1-Distill-Llama-70B (1024)	0.087	0.398	0.000	0.077	0.013	0.084
65	Qwen3.5-2B (1024)	0.000	0.312	0.000	0.033	0.006	0.051
66	Ministral-8B-Instruct-2410	0.109	0.228	0.000	0.023	0.004	0.047
67	Mistral-Small-3.1-24B-Instruct-2503	0.025	0.103	0.000	0.105	0.008	0.045
68	Mixtral-8x22B-Instruct-v0.1	0.171	0.084	0.000	0.040	0.004	0.037
69	Llama-3.2-1B-Instruct	0.000	0.190	0.000	0.024	0.006	0.033
70	Phi-3-mini-128k-instruct	0.027	0.094	0.000	0.024	0.003	0.022
71	Phi-3.5-MoE-instruct	0.000	0.070	0.000	0.023	0.004	0.017
72	Phi-4-mini-instruct	0.000	0.055	0.000	0.023	0.005	0.015
73	Mixtral-8x7B-Instruct-v0.1	0.000	0.016	0.000	0.040	0.003	0.013
74	Phi-3.5-mini-instruct	0.000	0.020	0.000	0.024	0.005	0.010
75	Phi-3-medium-128k-instruct	0.000	0.000	0.000	0.023	0.003	0.007
76	Qwen3.5-0.8B (1024)	0.000	0.000	0.000	0.023	0.003	0.007

Overview

This dataset contains agent game trajectories collected across multiple text-adventure game frameworks. Each trajectory records a full episode of an LLM-driven agent interacting with a text-based game environment, including observations, actions, scores, and (where available) thinking traces.

Frameworks

• alfworld
• jericho
• scienceworld
• textworld
• textworldexpress

Dataset Structure

Each row is a single trajectory with the following fields:

Field	Description
transcript_id	Unique identifier for the trajectory
framework	Game framework (e.g., textworld, jericho)
game	Specific game name
model	LLM used to drive the agent
agent_type	Agent architecture (e.g., zero-shot)
score	Final score achieved
max_score	Maximum possible score
normalized_score	Score normalized to [0, 1]
total_steps	Number of agent actions taken
turns	Full conversation history (JSON string)
score_progression	Score at each step (JSON string)

Note: thinking_available is true when thinking tokens were reported (total_thinking_tokens > 0), not when reasoning text is present. Some models report thinking tokens but return no reasoning text (e.g., gpt-5 and o3), and some rows include reasoning text in thinking without a reported count (e.g., Qwen/Qwen3-32B and deepseek-ai/DeepSeek-R1). For deepseek-ai/DeepSeek-R1-Distill-Llama-70B, total_thinking_tokens is estimated from the reasoning length (characters ÷ 4). API models that do not return their raw reasoning either return no reasoning text or a summary of it; for claude-opus-4.5, claude-opus-4.6, claude-sonnet-4.6, claude-opus-5.5 and gpt-6-astra, thinking holds this summary, not the raw reasoning.

Usage

from datasets import load_dataset

# Load a specific framework
ds = load_dataset("ccui46/tale_suite_trajectories", "alfworld")

# Load all frameworks
for fw in ["alfworld", "jericho", "scienceworld", "textworld", "textworldexpress"]:
    ds = load_dataset("ccui46/tale_suite_trajectories", fw)
    print(f"{fw}: {len(ds['train'])} trajectories")

Raw Data

Raw JSONL files (one line per trajectory, metadata flattened) are available under raw/ for direct download.

Rejected Actions

extras/rejected_actions.jsonl lists, one JSON object per line, the 259 actions that the harness of the 500-step gpt-6-astra and claude-opus-5.5 runs rejected. In the transcripts these steps show only a placeholder action: talesunsafeaction for malformed input or interpreter commands, and talesblockedcommand for blocked save, load, restore and undo commands.

Field	Description
transcript_id, model, game, seed	The trajectory, as in data/jericho
step, turn_index	The step, and the index of its agent turn in turns
action	The placeholder recorded in turns
proposed_action	What the model submitted (null in the two redacted games)
rejection_type	unsafe_action or blocked_state_command
rejection_reason	Why the harness rejected it (e.g., Frotz interpreter meta-command, empty or whitespace-only action)
action_source	Where the harness read the action from: tool for gpt-6-astra, unknown for claude-opus-5.5; null for blocked commands
sent_to_game	Whether the placeholder was sent to the game. It was only for the 6 blocked claude-opus-5.5 commands, which the game answered as an unknown verb; otherwise nothing reached the game and the step was wasted

In the gpt-6-astra runs, an empty action is the action tool's <ENTER> key, which is sent to the game as a bare Return.

Redacted Games

For two Jericho games, The Hitchhiker's Guide to the Galaxy (JerichoEnvHhgg) and Adventureland (JerichoEnvAdventureland), the observations and the model reasoning / reasoning summaries are redacted. In these 782 trajectories, every environment turn reads [observation withheld; regenerate with your own copy of the story file], and on agent turns raw_output holds only the action and thinking is null. Actions, seeds, scores and token counts are unchanged, so thinking_available and the token counts still describe the original run.

To regenerate the observations, load the game with your own copy of the story file and the transcript's seed, then send the actions (content of the agent turns) to the game engine in order.

If the authors or other interested parties would prefer that the observations and reasoning be redacted for other games, please contact us at textworld@microsoft.com or czcui@ucsd.edu.

Citation

@inproceedings{cui2026tales,
  title     = {{TALES}: Text Adventure Learning Environment Suite},
  author    = {Cui, Christopher Zhang and Yuan, Xingdi and Xiao, Ziang and Ammanabrolu, Prithviraj and C{\^o}t{\'e}, Marc-Alexandre},
  booktitle = {40th Conference on Neural Information Processing Systems (NeurIPS 2026), Track on Evaluations and Datasets},
  year      = {2026}
}

License

CC BY 4.0

Jericho Game Licenses

The release contains no game files. The license of each of the 54 Jericho games is listed below; for the two games whose observations and reasoning are withheld, see Redacted Games.

Open-source and freeware games

game	Title	License	Author	Year
JerichoEnvZork1	Zork I	MIT (Open Source)	Marc Blank and Dave Lebling	1980
JerichoEnvZork2	Zork II	MIT (Open Source)	Dave Lebling and Marc Blank	1981
JerichoEnvZork3	Zork III	MIT (Open Source)	Dave Lebling and Marc Blank	1982
JerichoEnvSpirit	SPIRITWRAK	GPL	D. S. Yu	1996
JerichoEnvYomomma	Raising the Flag on Mount Yo Momma	Open Source	Juhana Leinonen	2010
JerichoEnv905	9:05	Freeware	Adam Cadre	2000
JerichoEnvAcorncourt	The Acorn Court	Freeware	Todd S. Murchison	1997
JerichoEnvAdvent	Adventure	Freeware	William Crowther and Donald Woods	1976
JerichoEnvAfflicted	Afflicted	Freeware	Doug Egan	2008
JerichoEnvAnchor	Anchorhead	Freeware	Michael Gentry	1998
JerichoEnvAwaken	The Awakening	Freeware	Dennis Matheson	1998
JerichoEnvBalances	Balances	Freeware	Graham Nelson	1994
JerichoEnvCurses	Curses	Freeware	Graham Nelson	1993
JerichoEnvDeephome	Deephome: A Telleen Adventure	Freeware	Joshua Wise	1999
JerichoEnvDetective	Detective	Freeware	Matt Barringer	1993
JerichoEnvDragon	Dragon Adventure	Freeware	William Stott	2003
JerichoEnvEnter	The Enterprise Incidents	Freeware	Brian Desilets	2002
JerichoEnvGold	Goldilocks is a FOX!	Freeware	J. J. Guest	2002
JerichoEnvHuntdark	Hunter, in Darkness	Freeware	Andrew Plotkin	1999
JerichoEnvInhumane	Inhumane	Freeware	Andrew Plotkin	1984
JerichoEnvJewel	The Jewel of Knowledge	Freeware	Francesco Bova	1999
JerichoEnvKarn	Return to Karn	Freeware	Patrick Wigfull	1996
JerichoEnvLibrary	All Quiet on the Library Front	Freeware	Michael S. Phillips	1995
JerichoEnvLoose	Mother Loose	Freeware	Irene Callaci	1998
JerichoEnvLostpig	Lost Pig	Freeware	Admiral Jota	2007
JerichoEnvLudicorp	The Ludicorp Mystery	Freeware	Cal Henderson	2006
JerichoEnvMoonlit	The Moonlit Tower	Freeware	Yoon Ha Lee	2002
JerichoEnvMurdac	Monsters of Murdac	Former commercial (free)	Jonathan Partington	1982
JerichoEnvNight	Night at the Computer Center	Freeware	bonni mierzejewska	1996
JerichoEnvOmniquest	OMNIQuest	Freeware	Chris Barden and Chris Ethridge	1988
JerichoEnvPartyfoul	Party Foul	Freeware	Brooks Reeves	2010
JerichoEnvPentari	Pentari	Freeware	Howard A. Sherman	1998
JerichoEnvReverb	Reverberations	Freeware	Russell Wain Glasser	1996
JerichoEnvSnacktime	Snack Time!	Freeware	Hardy the Bulldog and Renee Choba	2008
JerichoEnvTemple	The Temple	Freeware	Johan Berntsson	2002
JerichoEnvTryst205	Tryst of Fate	Freeware	G. M. Zagurski	1997
JerichoEnvWeapon	The Weapon	Freeware	Sean Barrett	2001
JerichoEnvZenon	Escape from the Starship Zenon	Shareware	Andrew Gawthorpe	2000
JerichoEnvZtuu	Zork: The Undiscovered Underground	Freeware	Marc Blank, Michael Berlyn, and G. Kevin Wilson	1997

Commercial games

game	Title	Rights holder	Released
JerichoEnvAdventureland	Adventureland	Scott Adams	Observations and reasoning withheld
JerichoEnvBallyhoo	Ballyhoo	Infocom (Microsoft)	Full trajectories
JerichoEnvCutthroat	Cutthroats	Infocom (Microsoft)	Full trajectories
JerichoEnvEnchanter	Enchanter	Infocom (Microsoft)	Full trajectories
JerichoEnvHhgg	The Hitchhiker's Guide to the Galaxy	Infocom (Microsoft) / Adams Estate	Observations and reasoning withheld
JerichoEnvInfidel	Infidel	Infocom (Microsoft)	Full trajectories
JerichoEnvLurking	The Lurking Horror	Infocom (Microsoft)	Full trajectories
JerichoEnvPlanetfall	Planetfall	Infocom (Microsoft)	Full trajectories
JerichoEnvPlundered	Plundered Hearts	Infocom (Microsoft)	Full trajectories
JerichoEnvSeastalker	Seastalker	Infocom (Microsoft)	Full trajectories
JerichoEnvSherlock	Sherlock: The Riddle of the Crown Jewels	Infocom (Microsoft)	Full trajectories
JerichoEnvSorcerer	Sorcerer	Infocom (Microsoft)	Full trajectories
JerichoEnvSpellbrkr	Spellbreaker	Infocom (Microsoft)	Full trajectories
JerichoEnvTrinity	Trinity	Infocom (Microsoft)	Full trajectories
JerichoEnvWishbringer	Wishbringer	Infocom (Microsoft)	Full trajectories