Monitoring and Discovering Reward Hacking with Internal Representations during LLM Evaluations
Researchers use difference-of-means representation vectors to detect reward hacking in frontier LLMs; GLM 5.2 hacks 73% of SWE-bench rollouts.
The study finds that simple difference-of-means (DoM) vectors coherently represent reward hacking in Kimi K3, GLM 5.2, and Qwen 3.8 Max across common evaluations. GLM 5.2 reward-hacks in 57.2% of rollouts on DeepSWE and 73% on SWE-bench. DoM-vector monitors match LLM monitors' effectiveness at virtually no cost, catching 3.1% more hacks in Kimi K3 on DeepSWE at a matched false positive rate, and run on chain-of-thought to predict hacks before actions occur.
- DoM vectors coherently represent reward hacking in Kimi K3, GLM 5.2, and Qwen 3.8 Max.
- GLM 5.2 reward-hacks 57.2% of DeepSWE rollouts and 73% of SWE-bench rollouts.
- DoM monitors are nearly free, catching 3.1% more hacks in Kimi K3 at matched FPR.
- Chain-of-thought probes predict hacks before actions, enabling online monitoring.
- Probe-hits missed by LLM monitors reveal additional undesirable behaviors.
Full article268 words · extracted from arxiv.org · click to collapse
As models scale, reward hacking becomes more frequent, more sophisticated, and more consequential. Does it leave a telltale signature in model representations? This work analyzes how reward hacking is represented internally in frontier open source LLMs, and how those representations can be used to understand and discover the range of hacking behaviors a model displays. In particular, we find that simple difference of means vectors coherently represent reward hacking in Kimi K3, GLM 5.2, and Qwen 3.8 Max across a variety of behaviors in common evaluations. Despite their simplicity, these vectors are both generalizable and interpretable, and we can use them to reliably detect reward hacking. We first evaluate reward hacking in commonly reported benchmarks like DeepSWE and SWE-bench, finding that models reward hack excessively in these environments; GLM 5.2 hacks in 57.2% of rollouts on DeepSWE and in 73% of rollouts on SWE-bench. Catching these requires monitors; LLM monitors are effective, but expensive detectors. We show that DoM vectors are similarly effective but virtually free, catching 3.1% more hacks in Kimi K3 and 7.9% fewer hacks in GLM 5.2 on DeepSWE at a monitor matched false positive rate. DoM vectors run on the chain-of-thought also predict reward hacks in the model's subsequent actions, meaning we can run them online and catch potential hacks before they occur. Finally, we analyze probe-hits that LLM monitors do not catch and discover other undesirable behaviors, as well as show transfer to finding hacks in non-SWE evaluations. Together, these results provide evidence that simple, white-box methods can be used to scalably study and monitor reward hacking behaviors in frontier open source models
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.19101