EvoUndo: Recoverability-Constrained Self-Evolution for LLM Agent Harnesses
The paper finds that many useful agent self-modifications could not be safely undone without stronger recovery machinery.
In 600 unseen one-shot self-evolution tasks, the authors found 197 capability-improving mutations that failed recoverability checks. Standard repair methods recovered none of those failures under the original representation. Their extended recovery calculus raised oracle recovery to 191 of 197, with expressivity and exact state-address grounding identified as separate constraints. HF Daily Papers' note
In 600 unseen one-shot self-evolution tasks, the authors found 197 capability-improving mutations that failed recoverability checks. Standard repair methods recovered none of those failures under the original representation. Their extended recovery calculus raised oracle recovery to 191 of 197, with expressivity and exact state-address grounding identified as separate constraints. HF Daily Papers' note
score 5