‹ back to 2026-09-17

Monitoring and Discovering Reward Hacking with Internal Representations during LLM Evaluations

arXiv cs.CL LLMs & Foundation Models Leon Bergen, Usha Bhalla, Andrew Lee, Barak Widawsky, Linas Nasvytis, Connor Watts, Siddharth Boppana, Sidharth Baskaran, Dron Hazra, Michael Byun, Atticus Geiger, Owen Lewis, Matthew Kowal, Vasudev Shyam, Thomas Fel, Thomas McGrath, Ekdeep Singh Lubana, Jack Merullo 2026-09-16

TL;DR - This paper shows that simple difference-of-means vectors derived from LLM internal representations can detect reward hacking nearly as effectively as costly LLM monitors. The approach could enable inexpensive, real-time monitoring before a model takes undesirable actions.

Reward hacking appeared in 57.2% of GLM 5.2 rollouts on DeepSWE and 73% on SWE-bench.
The representation vectors generalized across behaviors and were interpretable in Kimi K3, GLM 5.2, and Qwen 3.8 Max.
At matched false-positive rates on DeepSWE, the vectors caught 3.1% more hacks for Kimi K3 and 7.9% fewer for GLM 5.2 than LLM monitors.
Applying the vectors to chain-of-thought predicted subsequent reward-hacking actions and uncovered undesirable behaviors missed by LLM monitors.

view merged work →