Is ReFT All We Needed?

ReFT: A Revolutionary Approach to Fine-tuning LLMs

ReFT (Representation Finetuning), introduced in Stanford's May 2024 paper, offers a groundbreaking method for efficiently fine-tuning large language models (LLMs). Its potential was immediately apparent, further highlighted by Oxen.ai's July 2024 experiment fine-tuning Llama3 (8B) on a single Nvidia A10 GPU in just 14 minutes.

Unlike existing Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA, which modify model weights or input, ReFT leverages the Distributed Interchange Intervention (DII) method. DII projects embeddings into a lower-dimensional subspace, enabling fine-tuning through this subspace.

This article first reviews popular PEFT algorithms (LoRA, Prompt Tuning, Prefix Tuning), then explains DII, before delving into ReFT and its experimental results.

Parameter-Efficient Fine-Tuning (PEFT) Techniques

Hugging Face provides a comprehensive overview of PEFT techniques. Let's briefly summarize key methods:

LoRA (Low-Rank Adaptation): Introduced in 2021, LoRA's simplicity and generalizability have made it a leading technique for fine-tuning LLMs and diffusion models. Instead of adjusting all layer weights, LoRA adds low-rank matrices, significantly reducing trainable parameters (often less than 0.3%), accelerating training and minimizing GPU memory usage.

Prompt Tuning: This method uses "soft prompts"—learnable task-specific embeddings—as prefixes, enabling efficient multi-task prediction without duplicating the model for each task.

Prefix Tuning (P-Tuning v2): Addressing limitations of prompt tuning at scale, Prefix Tuning adds trainable prompt embeddings to various layers, allowing task-specific learning at different levels.

LoRA's robustness and efficiency make it the most widely used PEFT method for LLMs. A detailed empirical comparison can be found in this paper.

Distributed Interchange Intervention (DII)

DII is rooted in causal abstraction, a framework using intervention between a high-level (causal) model and a low-level (neural network) model to assess alignment. DII projects both models into subspaces via orthogonal projections, creating an intervened model through rotation operations. A detailed visual example is available here.

The DII process can be mathematically represented as:

where R represents orthogonal projections, and the distributed alignment search (DAS) optimizes the subspace to maximize the probability of expected counterfactual outputs post-intervention.

ReFT – Representation Finetuning

ReFT intervenes in the model's hidden representation within a lower-dimensional space. The illustration below shows the intervention (phi) applied to layer L and position P:

LoReFT (Low-rank Linear Subspace Reft) introduces a learned projected source:

where h is the hidden representation, and Rs edits h in the low-dimensional space spanned by R. The LoReFT integration into a neural network layer is shown below:

During LLM fine-tuning, the LLM parameters remain frozen, and only the projection parameters (phi={R, W, b}) are trained.

Experimental Results

The original ReFT paper presents comparative experiments against full fine-tuning (FT), LoRA, and Prefix Tuning across various benchmarks. ReFT techniques consistently outperform existing methods, reducing parameters by at least 90% while achieving superior performance.

Discussion

ReFT's appeal stems from its superior performance with Llama-family models across diverse benchmarks and its grounding in causal abstraction, which aids model interpretability. ReFT demonstrates that a linear subspace distributed across neurons can effectively control numerous tasks, offering valuable insights into LLMs.

References

• Wu et al., ReFT: Representation Finetuning for Language Models
• Hu et al., LoRA: Low-Rank Adaptation of Large Language Models
• Zhuang et al., Time-Varying LoRA
• Liu et al., P-tuning v2
• Geiger et al., Finding alignments between interpretable causal variables and distributed neural representations
• Lester et al., The power of scale for parameter-efficient prompt tuning
• Pu et al., Empirical analysis of the strengths and weaknesses of Peft techniques for LLMs

(Note: Please replace the bracketed //m.sbmmt.com/link/6c11cb78b7bbb5c22d5f5271b5494381 placeholders with the actual links to the research papers.)

The above is the detailed content of Is ReFT All We Needed?. For more information, please follow other related articles on the PHP Chinese website!