Summary of Inform: Mitigating Reward Hacking in Rlhf Via Information-theoretic Reward Modeling, by Yuchun Miao et al.
InfoRM: Mitigating Reward Hacking in RLHF via Information-Theoretic Reward Modeling
by Yuchun Miao, Sen Zhang, Liang Ding, Rong Bao, Lefei Zhang, Dacheng Tao
First submitted to arxiv on: 14 Feb 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed framework, InfoRM, tackles the issue of reward misgeneralization in reinforcement learning from human feedback (RLHF) by introducing a variational information bottleneck objective to filter out irrelevant information. This approach identifies a correlation between overoptimization and outliers in the IB latent space, establishing it as a promising tool for detecting reward overoptimization. The Cluster Separation Index (CSI) is introduced to quantify deviations in the IB latent space, serving as an indicator of reward overoptimization. Extensive experiments demonstrate the effectiveness of InfoRM on various settings and RM scales. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary InfoRM helps solve the problem of reward misgeneralization in RLHF by using a special kind of math called information-theory. This approach finds patterns that are not important for humans, which can cause models to behave badly. The new framework also shows that when models are overoptimizing, they tend to go to strange places. We found that this happens because the model is paying attention to things that aren’t important. By identifying these unusual patterns, we can stop models from getting stuck in bad behaviors. |
Keywords
* Artificial intelligence * Attention * Latent space * Reinforcement learning from human feedback * Rlhf
