From Goals to Rewards: Self-Evolving Agents for Scientific Discovery

Autonomous agents are increasingly used to accelerate scientific discovery, yet most still depend on scientists to specify two things: which objectives to optimize, and how to score candidates against them. I present two frameworks that let an agent supply both. SAGA (Scientific Autonomous Goal-evolving Agent) addresses what to optimize. Since human-specified objectives are imperfect proxies that invite reward hacking, SAGA uses a bi-level design: an outer loop of LLM agents proposes new objectives and compiles them into scoring functions, while an inner loop optimizes candidates under the current ones—reaching over 90% success on tasks like antibiotic and materials design. DrugSAGE addresses how to build the predictive models behind those scoring functions. Rather than searching from scratch for every property, it accumulates and reuses cross-task experience—verified skills, evidence about what works, and recurring fixes—so strong reward models transfer. It ranks first among nine agents across 33 property prediction tasks and reaches a 0.935 normalized score on 17 held-out tasks with no test-time search. Together, they sketch an autonomous discovery loop where the agent decides what to pursue and equips itself to measure progress.