PATTERN Cited by 1 source

Bounded model mutation

Intent

Constrain model changes to minimal, scoped, validatable patches rather than full retraining — enabling rapid iteration while limiting regression risk.

Problem

Full adapter retraining is neither quick nor safe. On an 8B-parameter base model, training a high-rank adapter takes days, and every weight change risks regressing inputs that were already working. Research confirms both the opportunity and the danger:

• Meng et al. (NeurIPS 2022): factual behavior is partially localized; targeted edits can succeed without disturbing unrelated behavior
• Cohen et al. (TACL 2024): even precise edits produce ripple effects on logically related knowledge
• Pletenev et al. (2025): LoRA adapters absorb corrections reliably up to ~hundreds of examples, beyond which reasoning degrades

Mechanism

1. Scope the change: identify the specific bug / behavior to fix
2. Train minimal patch: LoRA rank <50, single GPU, <1 hour
3. Validate bounded impact: two-gate system (no regression + uncertainty flagging)
4. Deploy incrementally: canary with automatic rollback
5. Manage lifecycle: prevent patch accumulation via patterns/adapter-lifecycle-management

Trade-offs

Pro	Con
Same-day turnaround	Only works for localizable bugs
Low regression risk per patch	Stacked patches interact (CACE)
Cheap to train	Ceiling on total corrections per adapter stack
Easy to rollback	Requires deterministic evaluation to validate

Seen in

• sources/2026-07-14-airbnb-llm-evaluation-infrastructure — Airbnb's Layer 3: micro adapters as bounded mutations