Large Language Model Agents Are Not Always Faithful Self-Evolvers

This article presents a systematic investigation into the concept of experience faithfulness in self-evolving large language model (LLM) agents, analyzing their reliance on past experiences across 13 LLM backbones and 9 environments. The study reveals that while these agents depend on raw experience, they often misinterpret or disregard condensed forms of experience due to semantic limitations and internal biases, which raises concerns about their reliability in decision-making. These findings highlight the need for improved methods of integrating experience in LLMs, which is crucial for practitioners aiming to develop more robust self-evolving AI systems.