Contrastive Reasoning Alignment: Reinforcement Learning from Hidden Representations

TL;DR

CRAFT is a novel alignment framework that enhances model safety by aligning reasoning traces in hidden representations, improving robustness against jailbreak attacks.

Contribution

It introduces a contrastive reinforcement learning approach to align large reasoning models' hidden states for safety, outperforming existing defenses.

Findings

CRAFT achieves 79.0% improvement in reasoning safety.

CRAFT achieves 87.7% improvement in final-response safety.

Outperforms state-of-the-art defenses like IPO and SafeKey.

Abstract

We propose CRAFT, a red-teaming alignment framework that leverages model reasoning capabilities and hidden representations to improve robustness against jailbreak attacks. Unlike prior defenses that operate primarily at the output level, CRAFT aligns large reasoning models to generate safety-aware reasoning traces by explicitly optimizing objectives defined over the hidden state space. Methodologically, CRAFT integrates contrastive representation learning with reinforcement learning to separate safe and unsafe reasoning trajectories, yielding a latent-space geometry that supports robust, reasoning-level safety alignment. Theoretically, we show that incorporating latent-textual consistency into GRPO eliminates superficially aligned policies by ruling them out as local optima. Empirically, we evaluate CRAFT on multiple safety benchmarks using two strong reasoning models, Qwen3-4B-Thinking and R1-Distill-Llama-8B, where it consistently outperforms state-of-the-art defenses such as IPO and SafeKey. Notably, CRAFT delivers an average 79.0% improvement in reasoning safety and 87.7% improvement in final-response safety over the base models, demonstrating the effectiveness of hidden-space reasoning alignment.