PAG: Multi-Turn Reinforced LLM Self-Correction with Policy as Generative Verifier

TL;DR

PAG introduces a unified reinforcement learning framework enabling large language models to self-correct by selectively verifying and revising their outputs, improving reasoning accuracy without external verifiers.

Contribution

It proposes a novel verify-then-revise mechanism within a multi-turn RL paradigm, enhancing LLM self-correction and verification capabilities in a unified approach.

Findings

PAG improves self-correction accuracy across reasoning benchmarks.

Self-verification with PAG outperforms self-consistency methods.

Selective revision reduces unnecessary model revisions.

Abstract

Large Language Models (LLMs) have demonstrated impressive capabilities in complex reasoning tasks, yet they still struggle to reliably verify the correctness of their own outputs. Existing solutions to this verification challenge often depend on separate verifier models or require multi-stage self-correction training pipelines, which limit scalability. In this paper, we propose Policy as Generative Verifier (PAG), a simple and effective framework that empowers LLMs to self-correct by alternating between policy and verifier roles within a unified multi-turn reinforcement learning (RL) paradigm. Distinct from prior approaches that always generate a second attempt regardless of model confidence, PAG introduces a selective revision mechanism: the model revises its answer only when its own generative verification step detects an error. This verify-then-revise workflow not only alleviates model collapse but also jointly enhances both reasoning and verification abilities. Extensive experiments across diverse reasoning benchmarks highlight PAG's dual advancements: as a policy, it enhances direct generation and self-correction accuracy; as a verifier, its self-verification outperforms self-consistency.