New Protocol Enables Auditable Hypothesis Evolution for AI Scientists.

Researchers have proposed the Hypothesis Evolution Protocol (HEP), a new framework designed to make the scientific reasoning of AI agents more transparent and auditable. Current large language model (LLM) agents often bury their hypothesis generation, testing, and belief updates within unstructured logs, making it difficult for humans or even the agents themselves to track their scientific process. HEP formalizes these operations, providing explicit mechanisms for hypothesis generation, evaluation, and evolution. This allows LLM agents to systematically engage in a hypothesis-test-evidence-belief cycle, similar to human scientists. Tested on materials-science research tasks, HEP-equipped agents demonstrated improved generalization and more effective use of the protocol as the underlying LLM's capabilities increased, marking a significant step towards verifiable AI-driven scientific discovery.