CausalGame: Benchmarking Causal Thinking of LLM Agents in Games

2026-07-05

AI Summary

Current AI scientist agents lack effective evaluation of causal reasoning capabilities, particularly in confronting real-world scientific challenges such as selection bias, measurement error, and unobserved confounding. This work proposes the first interactive, gamified benchmark that integrates these three sources of confounding across 14 scientific discovery scenarios, requiring large language model (LLM) agents to actively design experiments, collect data, and submit solutions accompanied by causal explanations. The framework enables end-to-end assessment through an interactive environment, a causal reasoning scoring rubric, and controlled experimental protocols. Evaluated on 30 mainstream LLMs, the best-performing model achieved a survival rate of only 68.0% (against a theoretical upper bound of 78–85%), with merely 5–7% of dialogues receiving causal reasoning validation, highlighting a significant gap in current models’ capacity for scientific causal thinking.

Abstract

Building AI Scientist agents with Large Language Models (LLMs) has recently attracted growing attention. Since scientific discovery fundamentally relies on uncovering causal relationships from observations, the capability of causal thinking, i.e., distinguishing causation from correlation and recognizing hidden biases, is essential to LLM agents. Although a number of benchmarks exist for AI Scientists, none explicitly incorporate challenges from selection bias, measurement error, and hidden confounders that widely exist in real-world scientific discovery. To this end, we present CausalGame, a benchmark that evaluates the causal thinking capabilities of LLM agents through interactive games. CausalGame asks LLM agents to actively design experimental protocols, collect observation data, and derive a final solution with an explanation report. To emulate realistic scientific discovery challenges, we design 14 scenarios that incorporate selection bias, measurement error, and hidden confounders. Across 30 LLM agents, none demonstrates reliable causal thinking: the best model reaches only 68.0% survival against analytical optima of 78-85%, and merely 5-7% of sessions receive credits on the causal-reasoning rubrics. CausalGame provides a scalable and controlled testbed for evaluating the causal thinking of AI Scientist agents.

Problem

• causal thinking
• selection bias
• measurement error
• hidden confounders
• LLM agents

Innovation

• causal thinking
• selection bias
• hidden confounders
• measurement error
• interactive benchmark