Prover-Verifier Games improve legibility of LLM outputs

TL;DR

This paper introduces a training method inspired by Prover-Verifier Games to improve the legibility and checkability of LLM outputs, enhancing human verification and model alignment.

Contribution

The paper proposes a novel training algorithm that enhances LLM output legibility by training small verifiers and helpful/sneaky provers, improving robustness and transferability to human verification.

Findings

Helpful prover accuracy increases during training

Verifier robustness to adversarial attacks improves

Human verification accuracy improves with helpful prover solutions

Abstract

One way to increase confidence in the outputs of Large Language Models (LLMs) is to support them with reasoning that is clear and easy to check -- a property we call legibility. We study legibility in the context of solving grade-school math problems and show that optimizing chain-of-thought solutions only for answer correctness can make them less legible. To mitigate the loss in legibility, we propose a training algorithm inspired by Prover-Verifier Game from Anil et al. (2021). Our algorithm iteratively trains small verifiers to predict solution correctness, "helpful" provers to produce correct solutions that the verifier accepts, and "sneaky" provers to produce incorrect solutions that fool the verifier. We find that the helpful prover's accuracy and the verifier's robustness to adversarial attacks increase over the course of training. Furthermore, we show that legibility training transfers to time-constrained humans tasked with verifying solution correctness. Over course of LLM training human accuracy increases when checking the helpful prover's solutions, and decreases when checking the sneaky prover's solutions. Hence, training for checkability by small verifiers is a plausible technique for increasing output legibility. Our results suggest legibility training against small verifiers as a practical avenue for increasing legibility of large LLMs to humans, and thus could help with alignment of superhuman models.