Recoverability and Verifier Strength: Two Measured Preconditions for Unattended Agent Self-Improvement
- Published
- Source
- arXiv
- Paper number
- 1365
- Field
- AI / Agents
- arXiv ID
- 2610.27465
Key points
- The authors define two measured quantities, verifier strength and recoverability, and propose a rule that deploys a change without human sign-off only when both are adequate.
- The platform's LLM judge gave the same score regardless of input in 9 of 37 settings, failing silently with no error.
- Restoring a stored copy brought back the original behavior exactly, while a regenerated undo fell 0.16 below the restoration floor.
- A deployed change altered the outcome of 17% of tasks beyond the noise floor, and no rollback repairs that reach.
- The platform applying the rule allows prompt, tool and skill edits to run unattended and keeps weight updates under human sign-off.
- Controlled experiments with prompt edits on the τ 2-bench agent benchmark supported the rule's assumptions.
Paper links
External research summaries. These are not HDATF publications or measured product results.