Large Language Models to Diffusion Finetuning

TL;DR

This paper introduces a diffusion-based finetuning method for large language models that enhances test-time compute scalability, improves accuracy with more diffusion steps, and integrates guidance techniques for specialized tasks without altering original weights.

Contribution

The proposed method enables large language models to scale test-time compute via diffusion steps, preserving original weights and unifying autoregressive and diffusion approaches.

Findings

Monotonically increasing accuracy with more diffusion steps

Effective question answering on specific topics

Compatibility with traditional finetuning methods

Abstract

We propose a new finetuning method to provide pre-trained large language models (LMs) the ability to scale test-time compute through the diffusion framework. By increasing the number of diffusion steps, we show our finetuned models achieve monotonically increasing accuracy, directly translating to improved performance across downstream tasks. Furthermore, our finetuned models can expertly answer questions on specific topics by integrating powerful guidance techniques, and autonomously determine the compute required for a given problem by leveraging adaptive ODE solvers. Our method is universally applicable to any foundation model pre-trained with a cross-entropy loss and does not modify any of its original weights, fully preserving its strong single-step generation capabilities. We show our method is more effective and fully compatible with traditional finetuning approaches, introducing an orthogonal new direction to unify the strengths of the autoregressive and diffusion frameworks.