PATTERN Cited by 1 source
Adapter lifecycle management¶
Intent¶
Prevent accumulated LoRA adapter patches from degrading model quality through uncontrolled interaction, using three lifecycle rules that parallel software dependency management.
Problem¶
Stacked LoRA patches interact — a direct case of CACE ("changing anything changes everything") from Sculley et al. (2015). Each patch is self-contained in isolation, but the combined stack can drift in unpredictable ways as patches accumulate.
Three lifecycle rules¶
1. Fuse co-triggering patches¶
Patches firing on overlapping inputs share relational neighborhoods. Naive stacking creates subspace interference. Solution: learnable fusion resolves the interference (per Gao et al., 2024 on LoRA adapter composition).
2. Retrain on accumulation¶
When the number of patches against one category approaches the empirical ceiling (~hundreds of examples per Pletenev et al., 2025), fold the patches into a clean full retrain. This prevents the adapter stack from crossing the threshold where reasoning degrades and the model becomes overconfident in hard-to-detect ways.
3. Unload unused patches¶
Every loaded patch needs revalidation when upstream changes. Patches not triggered within a defined window are automatically unloaded, reducing: - Revalidation overhead - Interaction surface area - Inference-time compute cost
Analogy to software¶
| Adapter lifecycle rule | Software equivalent |
|---|---|
| Fuse co-triggering | Merge related PRs / consolidate patches |
| Retrain on accumulation | Major version release (clean slate) |
| Unload unused | Deprecate and remove dead code |
Seen in¶
- sources/2026-07-14-airbnb-llm-evaluation-infrastructure — Airbnb's three rules for managing the micro-adapter stack