HypoBench introduces a comprehensive benchmark to systematically evaluate how large language models generate scientific hypotheses across multiple disciplines.

• Assesses hypothesis quality through practical utility, generalizability, and discovery rate
• Spans 7 real-world tasks and 5 synthetic scenarios to test hypothesis generation capabilities
• Provides structured evaluation criteria for what makes a good scientific hypothesis
• Enables comparison of different LLM approaches to hypothesis formulation

For biological research, HypoBench offers a standardized way to assess AI-generated hypotheses that could accelerate scientific discovery, reduce experimental costs, and help researchers explore novel pathways in experimental biology.

HypoBench: Towards Systematic and Principled Benchmarking for Hypothesis Generation