Right in the Right Way: LM Training with Verifiable Rewards and Human Demonstrations

2026-07-01

AI Summary

This work addresses the limitations of existing reinforcement learning (RL) approaches that rely solely on verifiable rewards for language model training, which often fail to capture subjective human preferences regarding style and structure, leading to monotonous, unnatural outputs or reward hacking. The paper proposes the first RL framework that integrates adversarial learning with verifiable rewards: a discriminator learns unverifiable adversarial reward signals from human demonstrations and jointly optimizes them alongside task-specific rewards. This dual-objective approach effectively bridges the gap between supervised fine-tuning and RL, significantly outperforming baselines across code repair, story generation, and reward-hacking benchmarks—achieving lower edit distances in code fixes, higher diversity and stylistic fidelity in narratives, near-elimination of reward cheating, and consistently high task accuracy.

Abstract

RL with verifiable rewards (RLVR) has emerged as a powerful paradigm for training LMs on tasks with well-defined success metrics, such as code generation and mathematical reasoning. However, current RLVR methods optimize only what can be objectively scored, often neglecting subjective, non-verifiable aspects of human-like outputs, such as style and structure. This limitation leads to well-documented failure modes such as diversity collapse, unnatural-sounding responses, and reward hacking. We propose an adversarial generator-discriminator framework that augments verifiable rewards with a learned signal from human demonstrations. A generator model is trained using RL to maximize both task accuracy and an adversarial reward derived from a discriminator. The discriminator, trained alongside the generator policy, learns to distinguish human-written outputs from model-generated ones. The discriminator serves as a learned proxy for the human output distribution, providing feedback on aspects of generation that are difficult to formalize as scalar rewards. Across diverse domains, including bug fixing and open-ended generation, our approach consistently improves non-verifiable properties while preserving the accuracy gains of RLVR. In bug fixing, our method produces solutions with significantly lower edit distance compared to RLVR baselines while matching end performance. In story generation, our method significantly improves win rate while producing stories that are diverse and more human-like. And in a simple reward hacking benchmark, our method nearly eliminates model misbehavior while maintaining high benchmark scores. Together, these results show that our approach bridges RL and SFT, offering a scalable path toward jointly optimizing the verifiable and non-verifiable properties of a task.

Problem

• verifiable rewards
• human-like outputs
• reward hacking
• diversity collapse
• non-verifiable properties

Innovation

• RLVR
• adversarial training
• human demonstrations
• language model alignment
• non-verifiable rewards