Eureka: Human-Level Reward Design via Coding Large Language Models
What this work does
A frozen LLM proposes reward programs; inner-loop RL trains control policies and resulting task fitness/reflection guides reward-program evolution.
Connection to discovery
A template for learning or searching the training interface while keeping the final discovery evaluator externally specified.
Evidence
Inspected the nested reward-design objective, code-selection mechanism and reward-generation ablations.
Scope and limitations
The LLM designer is not itself updated by RL in this loop. Simulation control scores and reward novelty do not demonstrate scientific mechanism discovery.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.