Research map / Reinforcement learning / Policy gradient and actor-critic
Robust Optimization for Mitigating Reward Hacking with Correlated Proxies
Zixuan Liu, Xiaolin Sun, Zizhan Zheng
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train the discriminator more thoroughly (ORPO* vs ORPO) | ORPO · Traffic | worst-case reward: -1.96e+04 -> -1.35e+04 | Improved |
| train the discriminator more thoroughly (ORPO* vs ORPO) | ORPO · Traffic | true reward: 16.91 -> 10.26 | Hurt |
| train the discriminator more thoroughly (ORPO* vs ORPO) | ORPO · Pandemic | true reward: -1.04 -> 1.18 | Improved |
| replace ORPO with Max-Min policy optimization | ORPO · Traffic | worst-case reward: -1.96e+04 -> -268.31 | Improved |
| replace ORPO with Max-Min policy optimization | ORPO · Pandemic | worst-case reward: -5.31e+06 -> -63.29 | Improved |
| replace ORPO with Max-Min policy optimization | ORPO · Glucose | worst-case reward: -27.54 -> -1.71 | Improved |
| replace ORPO with Max-Min policy optimization | ORPO · RLHF | worst-case reward: -1.84 -> -0.10 | Improved |
| replace ORPO with Max-Min policy optimization | ORPO · RLHF | true reward: 8.30 -> 5.38 | Hurt |
| replace reward ensemble with Max-Min policy optimization | Ensemble · RLHF | worst-case reward: -1.70 -> -0.10 | Improved |
| replace reward ensemble with Max-Min policy optimization | Ensemble · RLHF | true reward: 2.31 -> 5.38 | Improved |
| add linear feature structure (Linear Max-Min vs Max-Min) | Max-Min · Traffic | linear worst-case reward: -0.06 -> 0.20 | Improved |
| add linear feature structure (Linear Max-Min vs Max-Min) | Max-Min · Pandemic | true reward: 1.25 -> 3.65 | Improved |
Closest papers by meaning
- Correlated Proxies: A New Definition and Improved Mitigation for Reward Hacking (ICLR 2025 · similarity 0.94)
- A Reduction Framework for Distributionally Robust Reinforcement Learning under Average Reward (ICML 2025 · similarity 0.88)
- Blending Imitation and Reinforcement Learning for Robust Policy Improvement (ICLR 2024 · similarity 0.87)
- Reward Hacking in Rubric-Based Reinforcement Learning (NeurIPS 2026 · similarity 0.87)
- Reward Certification for Policy Smoothed Reinforcement Learning (AAAI 2024 · similarity 0.87)
- Exact Policy Recovery in Offline RL with Both Heavy-Tailed Rewards and Data Corruption (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
