Fine-Tuning & Optimization · Emerging · Advanced

Representation Fine-Tuning

Also known as: ReFT, PyReFT

A parameter-efficient fine-tuning approach that learns interventions on hidden representations rather than modifying model weights.

What Representation Fine-Tuning is

ReFT operates on internal hidden states during forward passes, requiring up to 10x to 100x fewer parameters than LoRA.

How it works

Trains lightweight linear projection matrices attached to specific transformer layer activations.

Why it matters

Allows efficient multi-task personalization and lightweight adaptation for edge devices.

Common uses