PATTERN Cited by 1 source
Micro-adapter hotfix¶
Intent¶
Fix a specific production LLM bug in same-day turnaround, scoped to one issue, without full model retraining — analogous to a software hotfix but for model behavior.
Mechanism¶
A micro adapter is a small LoRA patch layered on top of an existing shared adapter:
- Rank < 50 — well below the few-hundred range used for full adapter retraining
- Training: <1 hour on a single GPU
- Scope: learns only the minimal correction for a specific bug
- Does not modify the underlying shared adapter weights
Deployment model¶
Ships like a software hotfix: 1. Scoped to one issue 2. Validated behind two gates: (a) no regression on expert-reviewed domains; (b) high-uncertainty outputs flagged for human review 3. Canary-deployed with automatic rollback 4. Same-day turnaround from bug identification to production deployment
Prerequisites¶
- Deterministic evaluation (Layer 2) — without reproducible measurement, same-day validation is impossible
- Evaluation noise decomposition — must confirm the bug is real, not judge/reference noise
Risks & mitigations¶
Per [Sculley et al., 2015] (CACE — "changing anything changes everything"), stacked patches interact. Mitigated by patterns/adapter-lifecycle-management: - Fuse co-triggering patches - Retrain on accumulation - Unload unused patches
Per [Pletenev et al., 2025], LoRA adapters degrade reasoning beyond ~hundreds of examples — micro adapters stay well below this ceiling.
Seen in¶
- sources/2026-07-14-airbnb-llm-evaluation-infrastructure — Airbnb's Layer 3: bounded, scoped model mutation with micro adapters