Monitoring Emergent Reward Hacking During Generation via Internal Activations
This work provides a complementary and earlier signal for detecting emergent misalignment in fine-tuned language models, improving post-deployment safety monitoring for developers and users of LLMs.
This paper addresses the problem of detecting reward-hacking behavior in fine-tuned large language models, which is difficult to identify from final outputs alone. The authors propose an activation-based monitoring approach using sparse autoencoders and linear classifiers on residual stream activations, demonstrating that internal activation patterns reliably distinguish reward-hacking from benign behavior across various models and generalize to unseen adapters.
Fine-tuned large language models can exhibit reward-hacking behavior arising from emergent misalignment, which is difficult to detect from final outputs alone. While prior work has studied reward hacking at the level of completed responses, it remains unclear whether such behavior can be identified during generation. We propose an activation-based monitoring approach that detects reward-hacking signals from internal representations as a model generates its response. Our method trains sparse autoencoders on residual stream activations and applies lightweight linear classifiers to produce token-level estimates of reward-hacking activity. Across multiple model families and fine-tuning mixtures, we find that internal activation patterns reliably distinguish reward-hacking from benign behavior, generalize to unseen mixed-policy adapters, and exhibit model-dependent temporal structure during chain-of-thought reasoning. Notably, reward-hacking signals often emerge early, persist throughout reasoning, and can be amplified by increased test-time compute in the form of chain-of-thought prompting under weakly specified reward objectives. These results suggest that internal activation monitoring provides a complementary and earlier signal of emergent misalignment than output-based evaluation, supporting more robust post-deployment safety monitoring for fine-tuned language models.