Are Self-Evolving Agents Continual Learners?
- Published
- Source
- arXiv
- Paper number
- 1370
- Field
- AI Agents
- arXiv ID
- 2610.25858
Key points
- The authors built four task streams that shift in topic, workflow, or environment, and tested ReasoningBank, GEPA, and Meta-Harness.
- On average, self-evolution raised scores, with terminal advantage of +3.7 for GEPA, +2.5 for ReasoningBank, and +1.4 for Meta-Harness.
- GEPA rewrote a single shared prompt, giving the highest plasticity but the largest average stability loss of −4.3.
- ReasoningBank localized updates through memory retrieval and achieved the best average stability of −1.9, at the cost of +3.3 plasticity.
- When similar queries required conflicting workflows, retrieval collision made 52% of BFCL-Miss episodes retrieve memories of the opposite workflow at the end of workflow drift.
- The authors conclude that current self-evolving agents remain incomplete continual learners and call for mechanisms that isolate task-specific updates and condition retrieval beyond surface similarity.
Paper links
External research summaries. These are not HDATF publications or measured product results.