Skip to content

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

Last updated · 571 distilled / 1,747 read