SilentEvo: How Proactive Agents Learn What to Check, Keep, and Change
Proactive agents watch more cases than they can handle. In every cohort they decide what to act on now, what to keep for later and what to let go, so remembering and forgetting are one decision. Forgetting is silent: a dropped case produces no label, so self-improvement that learns from observed failures cannot see it. We introduce silent evolution, in which an agent improves its own rules during deployment, taught only by audits drawn independently of its choices, so the audits include what it forgot. SilentEvo is, to our knowledge, the first agent that learns this way. It learns what not to forget at three timescales: cases (check, keep or let go), rules (edit, prune or hold) and evidence (how much audit history to keep), all scored by preventable harm. We prove that once verification misses a region, no amount of unverified data can settle a decision. The replays show the same limit. An editor that audits only the cases it kept never learns a rule it needs, in 96 of 96 streams. The same number of audits drawn independently finds that rule in 93 streams and saves 0.78 harm units per queue. Keeping cases by harm instead of recall cuts severe misses by 17%. Learning when to hold its rules cuts edits 3.5-fold while preventing more harm than the best tuned editor, and learned editing settings are confirmed on 128 locked fresh streams (p = 7 × 10^-12). On the public IMP-MARL simulator, SilentEvo certifies 84.0% of its selected inspections, against 47.6% for historical calibration.
Authors
- Xin Heng
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-12
- DOI
- https://doi.org/10.5281/zenodo.24157368
- Primary Topic
- AI-based Problem Solving and Planning
- Type
- preprint