DeepMind paper: self-evolving agents forget as they learn — fastest learner GEPA also breaks old tasks most
rohanpaul_ai · x · 2026-10-21
A new Google DeepMind/Mila paper (arXiv:2610.25858) asks whether self-evolving agents — which adapt by editing their own prompts, memory, and skill files with frozen weights — are effective continual learners.
- Built 4 drifting task streams from BIRD, BFCL, and ScienceWorld; evaluated ReasoningBank, GEPA, and Meta-Harness.
- Key finding: higher plasticity consistently costs stability. Self-modifying methods apply global edits to shared prompts/workspaces, interfering with unseen tasks and overwriting earlier knowledge; similarity-based retrieval localizes updates but retrieves foreign memories and suffers retrieval collisions.
- GEPA learned the most (+6.2 points/block) and lost the most elsewhere (-4.3). A +15.5-point gain from asking users for missing details fell to 2.5 after a later block rewrote that rule.
- Practical advice: keep task-specific rules separate and regression-test old tasks after every self-edit.
More from coding & agent
- Burkov: without explicit checklists, coding agents hand you 80%-done work — burkov · 2026-10-21
- Laude Institute launches Slingshots FOUR: 14 projects across 8 top institutions — a1zhang · 2026-10-21
- Research agent Épi sets 120 records, narrows n-queens constant interval 84x — dair_ai · 2026-10-21
- Compaction hooks in practice: keep last n messages verbatim, summarize the rest for full agent continuity — MoonL88537 · 2026-10-21
- The blank prompt box hides real operational burden: context, workflow and supervision fall on users — _AustinCalvert_ · 2026-10-21
- arXiv: Are Self-Evolving Agents Continual Learners? Plasticity comes at the cost of stability — rohanpaul_ai · 2026-10-21