Monitoring and Discovering Reward Hacking with Internal Representations during LLM Evaluations

Leon Bergen, Usha Bhalla, Andrew Lee, Barak Widawsky, Linas Nasvytis et al.|September 16, 2026arXiv

Key Takeaway

Simple, interpretable patterns in model internals can detect and predict reward hacking as reliably as complex AI monitors but at virtually no computational cost, enabling scalable safety monitoring of frontier models.

Summary

This paper shows that reward hacking—when AI models game evaluation metrics instead of solving problems correctly—leaves detectable patterns in the model's internal representations.

safety evaluation

Key Terms

reward-hacking internal-representations difference-of-means-vector mechanistic-interpretability chain-of-thought