K-Pool LoRA
Continual fine-tuning of frozen pre-trained LLMs via a K-snapshot adapter pool, frozen-encoder routing, and a sign-quantized optimizer.
Mechanism
A software-method for continual supervised fine-tuning of frozen pre-trained language models. The method maintains a fixed-size pool of LoRA adapter snapshots over a frozen base, selects the active slot per training batch and per inference query through a gradient-detached content-aware routing module, opens unfitted slots via a bootstrap-and-novelty rule with a deterministic claim window, and updates only the active-slot parameters via a sign-quantized momentum optimizer.
The independent claim recites a deliberate apparatus-to-software mapping with the Infinity provisional: K-pool weight retention, sign-quantized active update, regime-identified routing, and a fragility-aware retention policy on the adapter pool.
Prior-art differentiator
Mixture-of-LoRA-experts approaches (Lifelong-MoE, MILE, MINGLE, KeepLoRA, AdaMix) update routing modules through the language-modeling gradient. K-Pool LoRA's routing module is gradient-detached on a frozen sentence-transformer feature representation, with online statistical accumulation rather than gradient updates. The bootstrap-and-novelty slot-opening rule, the deterministic claim window preventing freshly-opened-slot winner-take-all collapse, and the sign-quantized optimizer applied to active-slot parameters only do not jointly appear in any single prior reference.
| Property | MoE-LoRA prior art | This filing |
|---|---|---|
| Router update path | LM gradient through router | Gradient-detached on frozen encoder |
| Slot-opening rule | Implicit / softmax-only | Bootstrap-and-novelty with deterministic claim window |
| Active-slot optimizer | Adam-class on full pool | Sign-quantized momentum, active slot only |
| Replay buffer | Often required | Not required |
Privacy and regulatory advantages
The structurally-replay-free property of the method extends from regulated-industry compliance (HIPAA, FINRA, ITAR) to consumer on-device personalization. Conventional continual-learning approaches that fight catastrophic forgetting through Experience Replay, gradient-episodic-memory, averaged-GEM, and maximally-interfered-retrieval all require retaining samples of previous training data on the device. K-Pool LoRA's K-snapshot adapter pool with frozen-encoder routing encodes learned behavior in adapter weights only; no raw training data is retained between fine-tuning sessions.
This is privacy-by-architecture rather than privacy-by-policy. GDPR Article 5(1)(c) data-minimization obligations, CCPA and CPRA right-to-delete provisions, EU AI Act high-risk-system audit requirements, China PIPL right-to-rectify, and India DPDP Act consent and storage-limitation rules all become architecturally trivial: there is no retained user data to delete, audit, rectify, or limit. Edge deployments on consumer devices (NPUs, mobile SoCs, AI PC silicon) inherit this property automatically; OEMs shipping K-Pool LoRA-based personalization do not need to architect bespoke replay-buffer governance for each jurisdiction.
The frozen-encoder routing module further eliminates the need to retain task identifiers or session metadata at inference time. The pool routes itself based on input embeddings; the inference path does not surface raw user state to any downstream observer. Replay-buffer storage cost (typically 1 to 10 gigabytes on a personal device after months of use) and the NPU cycles required for replay-pass training are both eliminated, with corresponding flash and battery savings.
Counsel posture
Twelve-month conversion and PCT deadline 2027-05-07. Track One Prioritized Examination concurrently with non-provisional recommended to secure first office action under the favorable Desjardins MPEP guidelines within four-to-six months. Specialized software-patent counsel familiar with PTAB Desjardins and Federal Circuit Recentive AI-eligibility frameworks targeted for non-provisional drafting (existing apparatus counsel was sufficient for Infinity but software-method drafting requires a litigator-prosecutor hybrid).