The authors formalize self-evolving agents as systems that update their harness (prompt, memory, skills) from task feedback while model weights remain frozen.
Self-evolving agents trade plasticity for stability in continual learning streams.
Evaluating three representative methods across drifting task streams reveals that self-modification and memory retrieval both accumulate interference over unseen and prior tasks.
Big Tech
Xiaoyin Chen · Yizheng Huang · Yinglun Xu · Ziyun Zhao · Emmett McQuinn · Julie Liu · +1 more
Mila – Quebec AI Institute · Université de Montréal · Google DeepMind
Research Digest··3 min read
The authors test whether nonparametric self-evolving LLM agents, which adapt by rewriting prompts, memory, or code without updating model weights, can function as continual learners.
§