Teaching Models to Verbalize Reward Hacking in Chain-of-Thought Reasoning

Abstract: LLMs trained with RL can engage in reward hacking--exploiting unintended strategies for high reward--without revealing this behavior in their chain-of-thought reasoning, making detection difficult and posing risks for high-stakes applications. We propose verbalization fine-tuning (VFT), a pre-RL intervention that trains models to explicitly acknowledge when they are influenced by prompt cues--hints which point to incorrect answers (e.g., "a Stanford professor thinks the answer is A"). To evaluate VFT, we subsequently train models with RL on environments where held-out prompt cues signal which incorrect answers will receive high reward, incentivizing models to reward hack by exploiting cues instead of reasoning correctly. We measure how often models exploit these cues without verbalizing it. After RL, only 6% of the VFT-trained model's responses consist of undetected reward hacks. In comparison, when we perform RL without VFT, the rate of undetected reward hacks goes up to 88%; with a debiasing baseline intervention, this increases further to 99%. VFT achieves this by substantially increasing how often models verbalize the influence of cues--from 8% to 42% after VFT, and up to 94% after RL--while baselines remain low even after RL (10% and 1%). Our results show that teaching models to explicitly verbalize reward hacking behavior before RL significantly improves their detection, offering a practical path toward more transparent and safe AI systems.