BRiTE: Bootstrapping Reinforced Thinking Process to Enhance Language Model Reasoning
This addresses the problem of unreliable reasoning in LLMs for tasks like math and coding, offering a novel method that is incremental in enhancing existing models.
The paper tackles the challenge of generating reliable reasoning processes in large language models by introducing BRiTE, a bootstrapping reinforced thinking process algorithm, which improves performance on math and coding benchmarks without human-annotated data, matching or exceeding supervised fine-tuning results.
Large Language Models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks, yet generating reliable reasoning processes remains a significant challenge. We present a unified probabilistic framework that formalizes LLM reasoning through a novel graphical model incorporating latent thinking processes and evaluation signals. Within this framework, we introduce the Bootstrapping Reinforced Thinking Process (BRiTE) algorithm, which works in two steps. First, it generates high-quality rationales by approximating the optimal thinking process through reinforcement learning, using a novel reward shaping mechanism. Second, it enhances the base LLM by maximizing the joint probability of rationale generation with respect to the model's parameters. Theoretically, we demonstrate BRiTE's convergence at a rate of $1/T$ with $T$ representing the number of iterations. Empirical evaluations on math and coding benchmarks demonstrate that our approach consistently improves performance across different base models without requiring human-annotated thinking processes. In addition, BRiTE demonstrates superior performance compared to existing algorithms that bootstrap thinking processes use alternative methods such as rejection sampling, and can even match or exceed the results achieved through supervised fine-tuning with human-annotated data.