Do LLMs Know a Good Hypothesis When They See One? Logit-Based Energy Scoring Outperforms Prompted LLM-as-Judge for Scientific Hypothesis Ranking

LLMs struggle with hypothesis evaluation when asked directly: When researchers prompted language models to rank scientific hypotheses (the "LLM-as-Judge" approach), they only correctly identified the true hypothesis 16.6% of the time—barely better than random guessing on a field of 16 options.

A model's internal confidence signals work better than its explanations: By measuring a language model's intrinsic confidence through "logit-based energy scoring" rather than asking it to explain its reasoning, researchers achieved 33% accuracy on average—double the performance of prompted ranking—suggesting models may "know" more than they can articulate.

Smaller, specialized models can outperform larger ones for this task: A 1-billion-parameter model using the confidence-based method reached 53.1% accuracy, outperforming larger models, indicating that size alone doesn't guarantee better scientific judgment and that evaluation method matters more than model scale.

Current AI evaluation methods risk favoring familiar ideas over novel ones: The paper identifies a critical bias in existing LLM-as-Judge approaches—they may systematically undervalue genuinely novel hypotheses in favor of conventional ones, which could stifle scientific innovation if these tools are deployed without safeguards.

Confidence-based evaluation deserves more research for trustworthy AI science: The findings suggest that measuring what models are confident about, rather than what they claim to believe, could be a more reliable foundation for AI-assisted scientific discovery workflows.