FALSIFYBENCH: Using Large Models to Play the 'Guess the Rule' Game to Test AI's Scientific Reasoning Ability

FALSIFYBENCH is an evaluation framework inspired by the classic Wason 2-4-6 task, designed to test the hypothesis-driven reasoning ability of large language models (LLMs). Key findings include: Reasoning-optimized models outperform instruction-tuned models; models actively seeking falsification are more successful; however, all models still have a significant gap from optimal performance. This framework provides a new perspective for evaluating the scientific reasoning ability of LLMs.

Large language models are being deployed as autonomous agents for scientific research, but traditional benchmark tests only focus on static question-answering and cannot capture the dynamic, iterative process of scientific inquiry. Inductive reasoning is the cornerstone of scientific thinking, involving hypothesis generation, evidence collection, and belief revision—parts missing from existing benchmarks.

FALSIFYBENCH simulates the scientific discovery process: models need to propose number triplets to test the hidden rule, and the system feedbacks whether they conform to the rule. The core steps of the task include hypothesis generation, evidence collection (designing experiments), and belief revision. This task reveals the common confirmation bias in humans—tending to verify hypotheses rather than look for counterexamples.

After evaluating 12 different LLMs, the findings are: 1) Reasoning models generally outperform instruction-tuned models; 2) Models actively seeking falsification perform significantly better (consistent with Popper's falsificationism); 3) All models are far from reaching optimal performance; 4) Typical failure modes include premature convergence, confirmation bias loops, and misinterpretation of feedback.

Implications of the research results for development: 1) Need to introduce interactive evaluation frameworks to replace static benchmarks; 2) Well-designed prompts can guide models to adopt effective reasoning strategies (e.g., requiring falsification); 3) In the short term, develop human-AI collaboration models (AI generates hypotheses, humans are responsible for falsification); 4) Training data needs to include more examples of falsification thinking.

Current limitations: FALSIFYBENCH is a simplified abstract task that does not cover complex scenarios in real scientific research (e.g., multimodal data, ambiguous feedback). Future directions: Expand to multimodal reasoning, testing on real scientific problems, and evaluating metacognitive abilities.

FALSIFYBENCH reveals the significant limitations of current LLMs in scientific reasoning. A model's text generation ability does not equal mature scientific reasoning ability; true scientific intelligence requires critical thinking (including self-criticism of hypotheses). This framework provides a roadmap for AI's development toward scientific intelligence.