New RACE framework enhances reinforcement learning with adaptive reward evolution

Researchers have developed a new framework called RACE (Reward-Adaptive Co-Evolution) to improve reinforcement learning in open-ended tasks. This framework addresses limitations of static rubrics and fixed co-evolution schedules by dynamically adapting to the informativeness of rewards. RACE classifies training data groups to update policies only with informative signals and switches between policy and rubric optimization phases based on reward saturation and improvement plateaus. The system also includes online threshold adaptation for dynamic adjustments. Experiments show RACE outperforms fixed-alternation baselines across multiple tasks and enhances rubric discrimination. AI

IMPACT This framework could improve the training efficiency and performance of AI agents in complex, open-ended tasks.

RANK_REASON The cluster contains a research paper detailing a new framework for reinforcement learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• Litmaps
• RACE
• ScienceCast
• scite Smart Citations

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →