Self-evolving LLM agents study: mid-tier models benefit most

A team of 17 researchers from multiple institutions published a paper analyzing the self-evolution capabilities of LLM agents that use editable external harnesses—prompts, skills, memories, and tools—to improve task execution. They distinguish two key capabilities: harness-updating (producing useful updates from execution evidence) and harness-benefit (improving task performance from updated harnesses). Contrary to expectations, harness-updating is surprisingly flat across model tiers: even a relatively small model like Qwen3.5-9B produced harness updates that led to gains comparable to those from Claude Opus 4.6. This suggests that the ability to generate useful harness modifications does not scale with base model capability, meaning weaker models can contribute valuable evolutionary improvements.

The more critical finding is that harness-benefit is non-monotonic with respect to base capability. Weak-tier models gained little from updated harnesses, mid-tier models benefited most, and strong-tier models actually benefited less than mid-tier. The researchers identified two failure modes in weak models: they either fail to activate the relevant harness artifacts (e.g., prompts or memory entries) or they activate them but then fail to follow instructions faithfully. These results imply that development efforts should prioritize the task-solving agent itself rather than the evolution mechanism. Additionally, training should focus on harness invocation and long-horizon instruction following, as these are the bottlenecks for weaker models. The full paper and source code are available on arXiv.