The Energy Loss Phenomenon in RLHF: A New Perspective on Mitigating Reward Hacking

Yuchun Miao, Sen Zhang, Liang Ding, Yuqi Zhang, Lefei Zhang, Dacheng Tao
Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:44076-44105, 2025.

Abstract

This work identifies the Energy Loss Phenomenon in Reinforcement Learning from Human Feedback (RLHF) and its connection to reward hacking. Specifically, energy loss in the final layer of a Large Language Model (LLM) gradually increases during the RL process, with an excessive increase in energy loss characterizing reward hacking. Beyond empirical analysis, we further provide a theoretical foundation by proving that, under mild conditions, the increased energy loss reduces the upper bound of contextual relevance in LLMs, which is a critical aspect of reward hacking as the reduced contextual relevance typically indicates overfitting to reward model-favored patterns in RL. To address this issue, we propose an Energy loss-aware PPO algorithm (EPPO) which penalizes the increase in energy loss in the LLM’s final layer during reward calculation to prevent excessive energy loss, thereby mitigating reward hacking. We theoretically show that EPPO can be conceptually interpreted as an entropy-regularized RL algorithm, which provides deeper insights into its effectiveness. Extensive experiments across various LLMs and tasks demonstrate the commonality of the energy loss phenomenon, as well as the effectiveness of EPPO in mitigating reward hacking and improving RLHF performance.

Cite this Paper

BibTeX
@InProceedings{pmlr-v267-miao25c, title = {The Energy Loss Phenomenon in {RLHF}: A New Perspective on Mitigating Reward Hacking}, author = {Miao, Yuchun and Zhang, Sen and Ding, Liang and Zhang, Yuqi and Zhang, Lefei and Tao, Dacheng}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {44076--44105}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/miao25c/miao25c.pdf}, url = {https://proceedings.mlr.press/v267/miao25c.html}, abstract = {This work identifies the Energy Loss Phenomenon in Reinforcement Learning from Human Feedback (RLHF) and its connection to reward hacking. Specifically, energy loss in the final layer of a Large Language Model (LLM) gradually increases during the RL process, with an excessive increase in energy loss characterizing reward hacking. Beyond empirical analysis, we further provide a theoretical foundation by proving that, under mild conditions, the increased energy loss reduces the upper bound of contextual relevance in LLMs, which is a critical aspect of reward hacking as the reduced contextual relevance typically indicates overfitting to reward model-favored patterns in RL. To address this issue, we propose an Energy loss-aware PPO algorithm (EPPO) which penalizes the increase in energy loss in the LLM’s final layer during reward calculation to prevent excessive energy loss, thereby mitigating reward hacking. We theoretically show that EPPO can be conceptually interpreted as an entropy-regularized RL algorithm, which provides deeper insights into its effectiveness. Extensive experiments across various LLMs and tasks demonstrate the commonality of the energy loss phenomenon, as well as the effectiveness of EPPO in mitigating reward hacking and improving RLHF performance.} }
Endnote
%0 Conference Paper %T The Energy Loss Phenomenon in RLHF: A New Perspective on Mitigating Reward Hacking %A Yuchun Miao %A Sen Zhang %A Liang Ding %A Yuqi Zhang %A Lefei Zhang %A Dacheng Tao %B Proceedings of the 42nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2025 %E Aarti Singh %E Maryam Fazel %E Daniel Hsu %E Simon Lacoste-Julien %E Felix Berkenkamp %E Tegan Maharaj %E Kiri Wagstaff %E Jerry Zhu %F pmlr-v267-miao25c %I PMLR %P 44076--44105 %U https://proceedings.mlr.press/v267/miao25c.html %V 267 %X This work identifies the Energy Loss Phenomenon in Reinforcement Learning from Human Feedback (RLHF) and its connection to reward hacking. Specifically, energy loss in the final layer of a Large Language Model (LLM) gradually increases during the RL process, with an excessive increase in energy loss characterizing reward hacking. Beyond empirical analysis, we further provide a theoretical foundation by proving that, under mild conditions, the increased energy loss reduces the upper bound of contextual relevance in LLMs, which is a critical aspect of reward hacking as the reduced contextual relevance typically indicates overfitting to reward model-favored patterns in RL. To address this issue, we propose an Energy loss-aware PPO algorithm (EPPO) which penalizes the increase in energy loss in the LLM’s final layer during reward calculation to prevent excessive energy loss, thereby mitigating reward hacking. We theoretically show that EPPO can be conceptually interpreted as an entropy-regularized RL algorithm, which provides deeper insights into its effectiveness. Extensive experiments across various LLMs and tasks demonstrate the commonality of the energy loss phenomenon, as well as the effectiveness of EPPO in mitigating reward hacking and improving RLHF performance.
APA
Miao, Y., Zhang, S., Ding, L., Zhang, Y., Zhang, L. & Tao, D.. (2025). The Energy Loss Phenomenon in RLHF: A New Perspective on Mitigating Reward Hacking. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:44076-44105 Available from https://proceedings.mlr.press/v267/miao25c.html.

Related Material