petri-bench: 699 causal-discovery episodes from LLM agents and classical baselines

Every episode is one attempt to find a hidden causal parameter in a procedurally generated simulation, under a fixed experiment budget. Nine frontier LLM agents and four classical experimental-design algorithms ran the same 30 tasks across five simulation engines.

Unlike answer-only evaluations, each episode is also audited for scientific method quality: whether the submitted conclusion was backed by an isolating experiment, whether that backing survives Holm correction for multiple comparisons, and whether the agent p-hacked by testing more parameters than it reported.

Headline result

Classical experimental-design algorithms substantially outperform every frontier LLM agent on these tasks.

solver	type	n	mean score	solved %	correct param %
adaptive	baseline	30	92.2	100.0	100.0
ofat	baseline	30	90.8	100.0	100.0
glm-5.1	llm	81	66.2	48.1	75.3
opus-4.8	llm	60	65.1	36.7	73.3
codex-gpt-5.5	llm	60	61.5	36.7	63.3
haiku-4.5	llm	90	60.7	35.6	66.7
codex-gpt-5.6-sol	llm	73	58.8	34.2	68.5
fable-5	llm	52	58.8	51.9	67.3
sonnet-4.6	llm	69	58.5	37.7	56.5
gemini-3.1-pro	llm	90	58.0	48.9	75.6
glm-5.2	llm	4	55.6	50.0	50.0
ofat-rand	baseline	30	40.5	33.3	33.3
random	baseline	30	19.0	3.3	20.0

adaptive and ofat are deterministic search strategies, not learned models. They find the right parameter every time within budget. No LLM agent exceeds a 52% solve rate. random submits without experimenting and marks the floor.

Read this before comparing audit columns

The audit pipeline did not run on every episode. For adaptive, ofat and random the audit fields are entirely absent, so audit_backing_survives_holm being false there means not audited, not failed. Do not read it as evidence that those strategies p-hack. ofat-rand was audited and is comparable.

Among LLM solvers, 21 of 579 episodes also lack an audit record (fable-5 10, gemini-3.1-pro 5, sonnet-4.6 3, and one each for glm-5.1, glm-5.2, opus-4.8). Rates below use audited episodes as the denominator, since an absent audit is not a failed one.

Score, solve rate and correct-parameter rate are comparable across all solvers.

Two further caveats: glm-5.2 has only 4 episodes and should not be ranked against the rest, and pilot-v0 (30 episodes) is a pre-registration pilot kept for provenance.

Method-quality findings among LLM agents

Detected p-hacking is rare but not zero: opus-4.8 in 3 episodes, and one episode each for codex-gpt-5.6-sol, gemini-3.1-pro and glm-5.1.

Holm survival, over audited episodes, runs from 80.0% (gemini-3.1-pro) to 93.9% (sonnet-4.6). It does not track raw score: gemini-3.1-pro and glm-5.1 have nearly identical correct-parameter rates (75.6% and 75.3%) but differ by 14 points on Holm survival. Separating method from answer is the point of the benchmark.

Schema

One row per episode.

field	meaning
episode_id, solver, solver_type	identity; type is llm, baseline or pilot
task_id, engine, seed, metric	task identity; engine is one of market, origin, morph, swarm, social
budget, replicates, candidate_params	experiment budget and the parameter search space
truth_param, truth_value, truth_effect	the hidden ground truth
correct_param, correct_direction, solved	outcome
calls_used, efficiency, within_budget	cost of the search
claim_validity, score, score_notes	scoring
audit_*	method audit: alpha, family size, distinct params tested, excess testing, survivors, submission backing, Holm survival, p-hacking flag
bench_version, commit, created_at	provenance
log_calls, log_claims, log_submission	full interaction trace, JSON-encoded strings

Engines

Five procedurally generated simulations: market (price dynamics), origin (replicator chemistry), morph (reaction-diffusion), swarm (collective motion), social (opinion dynamics). Tasks are generated from seeds, so the ground truth is known by construction rather than annotated.

Citation

@dataset{sozudogru_petribench,
  author    = {Sozudogru, Baris},
  title     = {petri-bench: causal-discovery episodes from LLM agents
               and classical baselines},
  doi       = {10.5281/zenodo.20618024},
  url       = {https://doi.org/10.5281/zenodo.20618024},
  publisher = {Zenodo}
}

Benchmark code is MIT licensed. This dataset is released under CC BY 4.0. ORCID: 0009-0002-6506-1720