reft-modifies-hidden-activations-not-weights-on-frozen-model
IN premise — summaries/2026/08/24/wiki-Fine-tuning_deep_learning.md
Created 2026-08-24T17:11:10+00:00
ReFT (Representation Fine-Tuning) modifies less than 1% of hidden representations (activations) rather than weights on a fully frozen base model, distinguishing it from weight-based PEFT methods like LoRA.
Summary
ReFT works by nudging a tiny sliver of the intermediate signals a frozen model produces at runtime, without changing a single parameter of the model itself. This makes it categorically different from LoRA, which adjusts the model's learned weights, so the two should not be lumped together as the same kind of fine-tuning trick.