04:00
2026-09-04
arxiv.org
artificial-intelligence
Where Does Harness-Optimization Value Live? Localized Gains and the Budget-Splitting Trap in Self-Evolving LLM Agents
A new arXiv preprint (2609.0289v1) introduces HARNESSEVO, which decomposes an LLM agent's textual harness into four separately evolvable slots, and finds that on ALFWorld with a frozen 7B backbone, ne…