模型微調／可控適應
Layer-Specific LoRA Experiments Show That the Best Update Location Differs for Learning Vocabulary, Facts, and Behavioral Policies
A new study restricted LoRA to the early, middle, or late blocks of Transformers and found that different learning objectives exhibit reproducible “adaptation geometries.” The findings elevate adapter placement from an implementation detail to a design variable for controlling knowledge transfer and side effects.
Yale University researchers released a controlled experiment on July 28 examining where LoRA should be inserted within a Transformer. Conventional fine-tuning places adapters in most or all layers, whereas the study updated only the early, middle, or late blocks and compared them with all-layer and parameter-matched configurations. The researchers measured whether models could acquire new content, transfer capabilities to unseen examples, and confine changes to the intended scope.
The tests covered five objectives: lexical binding, factual association, behavioral policy, causal mapping, and procedural reasoning. The results showed that there is no single optimal location across all objectives. Lexical binding was easier to acquire in early layers and produced fewer side effects there, although broader transfer still required more comprehensive updates. Among localized adapter configurations, factual association favored late layers. Behavioral learning separated “acquiring an action” from “activating a policy under specified conditions”: the former favored late layers, while the latter favored middle layers. Transfer on causal and procedural tasks was generally strongest with middle-layer or all-layer updates.
The authors call the curves for each objective across the three dimensions of acquisition, transfer, and boundedness its “adaptation geometry.” Most directional differences were reproduced across five model families and remained after matching parameter counts, suggesting that the results were not simply caused by one LoRA configuration using more weights. This is particularly important for enterprise fine-tuning: selecting rank, learning rate, and target modules using only a single validation set may yield an adapter that appears accurate but overgeneralizes or fails to transfer across different phrasings.
The code and data have been released publicly, allowing engineers to treat layer-specific adapter placement as an additional hyperparameter search dimension while building tests for acquisition, transfer, and boundedness. However, the paper used controlled and relatively narrow objectives, so its findings cannot yet be directly generalized to large-scale instruction fine-tuning, cross-lingual adaptation, or long-horizon agent training. The next step is to determine whether the same layer-specific patterns persist in larger models, Mixture-of-Experts (MoE) architectures, quantized LoRA, and configurations with multiple stacked adapters.