The authors begin by defining evolution views, each applying a different textual instruction to analyze the same accumulated experience and extract candidate knowledge.
Consistency-regularized co-evolution of knowledge and model improves LLM agents
Chen et al. demonstrate that seeking consistent utility gains across different analyses of experience yields more generalizable knowledge, boosting accuracy and efficiency.
Big Tech
Sirui Chen · Zhining Liu · Chao Zhang · Jingrui He · Zheng Li
University of Illinois Urbana-Champaign · Georgia Institute of Technology · Amazon
Research Digest··3 min read
Chen et al.
§