Safety Must Survive Self-Improvement: Why Failures Persist and How Agents Recover

简体中文

L1 mechanism diagnostic. Current validation and founder fallback; no paper editor or Amazon Bedrock reproduction.

Paper information

Field	Value
Paper	arXiv 2610.01073
Affiliations	First-author affiliation not listed on the paper's first page
First public release	2026-10-01 (arXiv v1)
Authors' code	No public author implementation found as of 2026-10-02.
Adapter / method	safe-self-improvement
Local code	src/auto_research/agent_research/latest_20261002.py

Original paper

Background and contribution

Separate detection, selection of today's runnable program, and source for the next edit. Revalidate all candidates in current conditions. If none passes, roll back to the founder instead of keeping an unsafe incumbent.

flowchart LR
  I[Current conditions] --> V[Revalidate candidates] --> S[Choose program or founder] --> N[Next edit source]

Key figure from the paper

Original paper diagram of validation and rollback. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism

\(p_{run}=argmax_{p\in V_{current}} score(p)\). When \(V_{current}=\varnothing\), run founder rollback.

Paper results

Historical scores retained unsafe programs in 22 of 48 histories. Full validation with validated rollback ended correct in all core trajectory-study cases while retaining over 43% deployment savings.

Local reproduction

Compare disabled/default versus enabled. Three-seed results are diagnostic_only=true; transitions, gradients, and invariants do not establish paper-scale effects.

Reproduction limits

Validation and founder fallback only; no paper editing model or Amazon Bedrock experiment.