skumar-ml/belief-world-models

BB-WM agent trajectories

Trajectories for Towards a Belief-Based World Model for LLM Agents (BB-WM).

ReAct and ReflAct agents are evaluated on ALFWorld, ScienceWorld, and BabyAI-Text. The world-model condition is the only thing that changes between variants. Temperature is 0. Qwen3-14B and Sonnet 4.6 are run with thinking off.

Code and configs: the BB-WM repo (eval_baselines.py, configs/task/, configs/model/).

Layout

<benchmark>/<agent>/<permutation>/<model>/<split>[/runN]/

Path field	Values	Where it comes from
benchmark	alfworld, sciworld, babyai	benchmark in the task config
agent	react, reflact	agent in the task config
permutation	see the table below	permutation in the task config. This is the world-model condition.
model	Llama-3.1-8B-Instruct, Qwen3-14B, claude-sonnet-4-6	model_tag in the model config, otherwise the last segment of model_name
split	unseen, unseen_zipf	--split test is written as unseen. ALFWorld --placement zipf is written as unseen_zipf.
runN	run1, run2, run3	--run N. Omitted --run writes the episodes directly into the split directory.

--method is a metrics label only. It does not change the path. The task config does.

Example: alfworld/reflact/wm_walle/Qwen3-14B/unseen/run2/ is ReflAct + BB-WM, Qwen3-14B, ALFWorld test split, seed 2. That is --exp_config alfworld_reflact_walle_wm --agent_config openrouter_qwen3_14b (or qwen3_14b) --split test --run 2.

Permutation to experiment config

The same paper condition uses a different folder name and --exp_config on each benchmark. react and reflact in the path select the row.

Folder	Paper condition	ALFWorld --exp_config	ScienceWorld --exp_config	BabyAI-Text --exp_config
original (ALFWorld, BabyAI) or regular (ScienceWorld)	Original, no world model	alfworld, alfworld_reflact	sciworld_react, sciworld_reflact	babyai_react, babyai_reflact
wm	Belief	alfworld_react_wm, alfworld_reflact_wm	sciworld_react_wm, sciworld_reflact_wm	babyai_react_wm, babyai_reflact_wm
walle (ALFWorld) or walle_oracle (ScienceWorld, BabyAI)	WALL-E	alfworld_react_walle, alfworld_reflact_walle	sciworld_react_walle_oracle, sciworld_reflact_walle_oracle	babyai_react_walle_oracle, babyai_reflact_walle_oracle
wm_walle (ALFWorld) or walle_oracle_wm (ScienceWorld, BabyAI)	BB-WM	alfworld_react_walle_wm, alfworld_reflact_walle_wm	sciworld_react_walle_oracle_wm, sciworld_reflact_walle_oracle_wm	babyai_react_walle_oracle_wm, babyai_reflact_walle_oracle_wm
wm_det	Memory	alfworld_react_wm_det, alfworld_reflact_wm_det	sciworld_react_wm_det, sciworld_reflact_wm_det	—
wm_no_prob	Belief-NoProb	alfworld_react_wm_no_prob, alfworld_reflact_wm_no_prob	—	—

In each cell the first config is ReAct and the second is ReflAct. alfworld with no suffix is ReAct + Original.

Environment classes inside those configs:

Folder	ALFWorld env_class	ScienceWorld env_class	BabyAI-Text env_class
original / regular	AlfWorldEnv	SciWorldEnv	BabyAIEnv
wm	WMAlfWorldEnv	WMSciWorldEnv	WMBabyAIEnv
walle / walle_oracle	WalleAlfWorldEnv	WalleOracleSciWorldEnv	WalleOracleBabyAIEnv
wm_walle / walle_oracle_wm	WalleWMAlfWorldEnv	WalleWMOracleSciWorldEnv	WalleWMOracleBabyAIEnv
wm_det	DeterministicWMAlfWorldEnv	DeterministicWMSciWorldEnv	—
wm_no_prob	NoProbWMAlfWorldEnv	—	—

ALFWorld WALL-E uses the released neurosymbolic rules. ScienceWorld and BabyAI-Text WALL-E use an oracle validity check against the simulator.

Paper figures use these folders:

• Main grid: original or regular, wm, walle or walle_oracle, wm_walle or walle_oracle_wm, split unseen.
• Memory vs. Belief: original or regular, wm, wm_det.
• Zipf ALFWorld priors: split unseen_zipf, folders original, wm, wm_no_prob.

Model folder to model config

Both local vLLM and OpenRouter writes use the same folder name.

Folder	--agent_config	model_name
Llama-3.1-8B-Instruct	react_llama8b	meta-llama/Llama-3.1-8B-Instruct
Llama-3.1-8B-Instruct	openrouter_llama8b	meta-llama/llama-3.1-8b-instruct
Qwen3-14B	qwen3_14b	Qwen/Qwen3-14B
Qwen3-14B	openrouter_qwen3_14b	qwen/qwen3-14b
claude-sonnet-4-6	litellm_claude_sonnet	aws/claude-sonnet-4-6

run_config.json records which of these configs produced the directory. API keys in that file are redacted.

Splits and task ids

Benchmark	Split folder	Episodes	Step budget	Task id in the filename
ALFWorld	unseen	134	30 env steps	integer game index, 0.json …
ALFWorld	unseen_zipf	134	30 env steps	same indices, Zipf(α=1.25) placement over valid receptacles
ScienceWorld	unseen	211	per-task budget (the harness default in the config is 100)	task-1-boil_21.json and similar
BabyAI-Text	unseen	100	64 env steps	<family>_<index>.json

BabyAI-Text is zero-shot on BabyAI-MixedTrainLocal-v0 with num_dists=0. These runs keep the families goto, pickup, putnext, and pick_up_seq_go_to. Belief and BB-WM use a uniform prior over the 6×6 walkable interior once the room is localized.

unseen_zipf is present for original, wm, and wm_no_prob.

Files in each run directory

File	Contents
<task_id>.json	One episode. A JSON array. Every entry except the last is a chat turn, {"role", "content"}. The last entry is metadata.
metrics.json	Aggregates for that directory: success, reward, steps, query counts, WALL-E rejections.
run_config.json	Task config, model config, and CLI args for the run.
log.txt	Harness log.

Episode metadata (last array element):

Field	Meaning
success	Goal achieved.
reward	Episode return. Boolean-like on ALFWorld and BabyAI-Text. ScienceWorld score on ScienceWorld.
finished	Episode ended.
terminate_reason	Why it ended (success, step limit, and similar).
steps	Environment steps. World-model queries and WALL-E rejections are not included.
query_steps	World-model queries.
walle_rejections	Actions rejected by WALL-E.
gated_steps	No-ops from the field-of-view gate.
max_steps	Per-episode budget when the env recorded one.
reward_trace	ScienceWorld running-max score after each env step. Empty on ALFWorld, BabyAI-Text, and older dumps.
error	Set when the episode recorded an error.

metrics.json also stores method, model, split, permutation, n_tasks, and max_steps. On BabyAI Belief and BB-WM, method (and sometimes permutation) can read wm_prior or walle_oracle_wm_prior. That string is the CLI label. The path and run_config.json → cli_args.exp_config are the condition names in the table above.

Which directory is one evaluation

Every directory that contains <task_id>.json files is one evaluation.

Llama-3.1-8B-Instruct and Qwen3-14B main-grid runs are run1, run2, and run3. Sonnet 4.6 is run1 on ALFWorld and BabyAI-Text. On ScienceWorld it is run1, run2, and run3.

wm_det, wm_no_prob, and every unseen_zipf run are run1 only.

alfworld/react/original/Qwen3-14B is not in this release (those files were lost).

Citation

Shubham Kumar, Harshit Kumar, Narendra Ahuja, Saurabh Jha.
Towards a Belief-Based World Model for LLM Agents.
arXiv:2609.00455.