Xiaoyin Chen et al. (7 authors) · 2026-10-19
Are Self-Evolving Agents Continual Learners?
“Evaluating three representative methods spanning experience retrieval and self-modification (ReasoningBank, GEPA, and Meta-Harness), we find that although self-evolution improves aggregate performance on most streams, higher plasticity consistently comes at the cost of stability.”
- Who
- Xiaoyin Chen et al. (7 authors)
- Date
- 2026-10-19
- Evidence
- Low certainty Starts at High: measured by someone other than the developer.
Very lowLowModerateHigh
High ↓ indirect measure ↓ imprecise → Low
- ↓ indirect measure. Three methods on one base model, with streams of 100 to 200 tasks, stand in for current self-evolving agents in general. “We hold the base model fixed across methods and streams so that differences arise from the evolver alone, and use Gemini 3.6 Flash”
- ↓ imprecise. Each hold-out set has 50 tasks per group (90 for ScienceWorld). Many per-stream gaps between methods are 1 or 2 points. The paper does not say how many evolution runs each number comes from and reports no spread across runs. “Every stream consists of consecutive blocks of 50 training tasks per group, evaluated on disjoint hold-out sets”
- Replication
- No rerun found. The paper links no code. We found no rerun by another group. Searched up to 2026-10-28. all claims
- Checked
- 2026-10-28, against the source page
- Links
- Source Paper
Also on the page
- The authors built four task streams from BIRD, BFCL and ScienceWorld. Each stream trains one task group at a time in blocks of 50 tasks, and the group changes in topic, in workflow, in environment, or in all three. The agent is not told when the group changes.
- ReasoningBank retrieves the most similar stored memory item. GEPA rewrites a strategy section of the system prompt. Meta-Harness edits the system prompt, memory and skill files, and controller code. Model weights stay frozen.
- Averaged over the four streams, GEPA had the highest plasticity (+6.2 points) and the largest stability loss (-4.3). ReasoningBank had the best stability (-1.9) with plasticity of +3.3. Meta-Harness looked stable on topic and workflow drift "only because its validation gate rejects half or more of the candidate updates".
- All runs use one base model, Gemini 3.6 Flash. On ReasoningBank's retrievals of unrelated memories, the authors say a similarity threshold could help, but "query similarity alone struggles to distinguish how related two queries actually are": queries on different topics in the same text-to-SQL task match at 0.60 to 0.64, against 0.56 to 0.63 across different tasks and domains.