Online Learnability of Chain-of-Thought Verifiers: Soundness and Completeness Trade-offs

TL;DR

This paper introduces an online learning framework for chain-of-thought verifiers in large language models, addressing the trade-offs between soundness and completeness to improve reasoning accuracy and safety.

Contribution

It develops novel theoretical bounds and algorithms for learning verifiers that balance error types, enhancing reasoning verification and generator performance.

Findings

Proposes a new online learning framework for verifiers.

Provides optimal algorithms for mistake trade-offs.

Demonstrates improved reasoning accuracy with learned verifiers.

Abstract

Large Language Models (LLMs) with chain-of-thought generation have demonstrated great potential for solving complex reasoning and planning tasks. However, the output of current LLMs is not fully reliable and needs careful verification. Even if LLMs get more accurate over time, learned verifiers can help increase trust, enforce safety constraints, and ensure alignment with personal preferences. A major challenge in learning verifiers, however, especially when their output will be used by the generator to improve its reasoning, is that the feedback loop between generator and verifier may produce substantial distribution shift. Motivated by this challenge, we propose an online learning framework for learning chain-of-thought verifiers that, given a problem and a sequence of reasoning steps, check the correctness of the solution. Highlighting the asymmetric role of soundness errors (failure in catching errors in a reasoning trace) and completeness errors (flagging correct reasoning steps as wrong), we introduce novel extensions of the Littlestone dimension which tightly characterize the mistake bounds for learning a verifier in the realizable setting. We provide optimal algorithms for finding the Pareto-frontier (the smallest total number of mistakes given a budget of soundness mistakes) as well as for minimizing a linear combination of asymmetric costs. We further show how our learned verifiers can be used to boost the accuracy of a collection of weak generators, and enable generation of proofs beyond what they were initially trained on. With the mild assumption that one of the generators can generate the next reasoning step correctly with some minimal probability, we show how to learn a strong generator with small error and abstention rates.