MMODELYST
Papers/Directional Alignment Mitigates Reward Hacking in Reinforcement Learning for Language Models
PAP

Directional Alignment Mitigates Reward Hacking in Reinforcement Learning for Language Models

May 24, 2026

arXiv
Abstract

Reward hacking arises when a model improves a proxy reward by exploiting shortcuts rather than solving the intended task. We study this failure mode through the geometry of reinforcement learning updates in language models and argue that hacking emerges when optimization drifts away from a stable low-dimensional learning trajectory. We analyze this drift through dominant singular directions of parameter updates and show that reward-hacking runs exhibit substantially larger directional change than clean runs. Motivated by this observation, we introduce trusted-direction projection, which constrains gradients to remain within a clean reference subspace. Across reward-hacking experiments on mathematical reasoning, the proposed approach delays shortcut exploitation and better preserves task performance.

Select text to highlight · click a highlight to remove · saved in this browser only
Authors
Wenlong Deng, Jiaji Huang, Kaan Ozkara, Yushu Li, Christos Thrampoulidis, Xiaoxiao Li, Youngsuk Park
Your notes (browser-local)
saved
arXiv:2605.25189