Reproduce “Are Self-Evolving Agents Continual Learners?” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiaoyin Chen, Yizheng Huang, Yinglun Xu, Ziyun Zhao, Emmett McQuinn, Julie Liu, Hope Yao
The authors test whether self-evolving agents can learn task streams while retaining earlier skills, using three harness-update methods around a frozen model. GEPA and Meta-Harness change shared harness components, while ReasoningBank retrieves stored experiences. Across four constructed streams, aggregate performance improves in most settings, but greater plasticity is associated with lower stability; topic drift yields near-zero or negative gains.
Key insights
- Self-evolution can improve average performance while still degrading performance on previously learned or not-yet-seen task groups; stronger within-group learning was associated with larger stability losses.
- Shared instructions can generalize a task-specific rule too broadly, while retrieval can surface a memory that is topically similar but prescribes the wrong workflow.
- Topic similarity alone is insufficient to decide whether an experience applies: BFCL-Miss and BFCL-Base had highly similar query embeddings despite requiring conflicting clarification behavior.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.