EvoUndo Finds 197 of 600 Self-Evolution Mutations Improve Capability but Cannot Be Undone, and Prompting Repairs Zero of Them
EvoUndo (arXiv 2608.28363, 2026-08-28, cs.AI) tests whether an agent's runtime edits to its own prompts, tools, middleware and harness can be reversed in states other than the one where they were made. Across 600 unseen one-shot self-evolution tasks it identifies 197 capability-improving mutations that fail recoverability verification, and conventional repair strategies recover 0 of 197 under the original recovery language. A protocol-locked 2x2 intervention separates the two bottlenecks: exact state-address grounding lifts recovery from 0/48 to 38/48 (79.2%) where the original language suffices, while extending the recovery calculus reaches 142/143 (99.3%) in the oracle-defined stratum, with a model-dependent negative interaction on gpt-oss-120b (133/143) that a Qwen3.8-27B replication does not reproduce.
↳ Follow the thread