CausalGame: Benchmarking Causal Thinking of LLM Agents in Games

TL;DR

CausalGame is a benchmark for the causal-thinking capability of LLM agents cast as AI Scientists: instead of static QA, the agent plays 14 interactive game scenarios that require designing an experimental protocol, collecting observations, and producing an explanation report. Scenarios explicitly inject the three failure modes of real-world scientific discovery — selection bias, measurement error, and hidden confounders — that existing AI-Scientist benchmarks omit. Across 30 evaluated LLM agents, none demonstrates reliable causal thinking: the best-performing agent reaches only 68.0% survival against analytical optima of 78–85%, and only 5–7% of sessions earn credit on the causal-reasoning rubrics. The benchmark provides a controlled testbed for measuring the gap between correlation-fitting and genuine causal inference in autonomous research agents.

Key claims

Existing AI-Scientist benchmarks omit the three canonical statistical challenges of scientific discovery — selection bias, measurement error, and hidden confounders — and CausalGame is designed as an interactive-game benchmark that explicitly incorporates all three across 14 scenarios [§Abstract].
Evaluation is agentic and multi-stage: the LLM must actively design an experimental protocol, collect observation data through interaction, and derive a final solution accompanied by an explanation report — not answer a fixed-form question [§Abstract].
Across 30 LLM agents evaluated, the best model reaches only 68.0% survival against an analytical optimum band of 78–85%, indicating a persistent gap between frontier LLM performance and provably optimal causal-inference play [§Abstract].
Only 5–7% of agent sessions receive credit on the causal-reasoning rubrics — the outcome-level performance ceiling is not matched by any observable causal-reasoning process in the agent’s trajectory [§Abstract].
The benchmark positions itself as scalable and controlled: because each scenario is a game with known ground-truth causal structure, the analytical optimum and the rubric grades are computable rather than crowd-sourced [§Abstract].

Method

CausalGame packages 14 game scenarios in which an LLM agent, prompted as an AI Scientist, must uncover a hidden causal structure through active experimentation. Each scenario embeds one or more of the three real-world confounds — selection bias, measurement error, hidden confounders — so a purely correlational strategy is provably suboptimal. The agent’s session proceeds in three phases: protocol design (what to observe and under what interventions), interactive data collection (rolling out the protocol against the game engine), and final report generation (a solution plus an explanation of the causal chain). Grading has two axes: an outcome-level “survival” score against the game’s analytical optimum, and rubric-level credit on whether the explanation report actually names the correct causal-reasoning move (identifying the confounder, correcting the bias, adjusting for the measurement error).

Results

Best agent survival: 68.0% versus analytical optima of 78–85% across the 14 scenarios [§Abstract].
Causal-reasoning rubric credit: only 5–7% of sessions [§Abstract] — the reporting stage is where all evaluated models fail, even when their outcome scores are non-trivial.
30 LLM agents evaluated (specific model list not enumerated in the abstract).

Why it’s interesting

CausalGame slots directly into the open questions of the AI-for-AI Research concept page: the cluster’s filed systems (Kosmos, FARS, ASI-ARCH, Karpathy autoresearch) all run closed-loop experiment → analysis → next-experiment loops on GPU clusters, but none of the filed evaluations isolate whether the loop’s reasoning is actually causal or just correlation-fitting at scale. Kosmos’s 57.9% expert-graded accuracy on cross-domain synthesis statements (Kosmos: An AI Scientist for Autonomous Discovery §2.1) hints at this gap without measuring it directly; CausalGame’s 5–7% rubric-credit rate is the first filed quantitative measurement of the causal-reasoning bottleneck. It also complements the env-as-package pattern tracked in RL Environment Platforms: unlike Toolathlon-GYM or SETA which measure end-to-end task completion, CausalGame’s rubric is a process grade over the agent’s causal explanations. Adjacent to Agent² RL-Bench: Can LLM Agents Engineer Agentic RL Post-Training? which benchmarks whether agents can engineer RL pipelines — CausalGame benchmarks the earlier upstream question of whether they can reason about experimental design at all.