AI Paper Atlas

Research map / Reinforcement learning / Policy gradient and actor-critic

Robust Optimization for Mitigating Reward Hacking with Correlated Proxies

Zixuan Liu, Xiaolin Sun, Zizhan Zheng

ICLR 2026 · Reinforcement learning › Policy gradient and actor-critic · approach: Model-based RL with guarantees · 1 citations (OpenAlex)

Read the paper

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 changedWhereMeasuredEffect
train the discriminator more thoroughly (ORPO* vs ORPO)ORPO · Trafficworst-case reward: -1.96e+04 -> -1.35e+04Improved
train the discriminator more thoroughly (ORPO* vs ORPO)ORPO · Traffictrue reward: 16.91 -> 10.26Hurt
train the discriminator more thoroughly (ORPO* vs ORPO)ORPO · Pandemictrue reward: -1.04 -> 1.18Improved
replace ORPO with Max-Min policy optimizationORPO · Trafficworst-case reward: -1.96e+04 -> -268.31Improved
replace ORPO with Max-Min policy optimizationORPO · Pandemicworst-case reward: -5.31e+06 -> -63.29Improved
replace ORPO with Max-Min policy optimizationORPO · Glucoseworst-case reward: -27.54 -> -1.71Improved
replace ORPO with Max-Min policy optimizationORPO · RLHFworst-case reward: -1.84 -> -0.10Improved
replace ORPO with Max-Min policy optimizationORPO · RLHFtrue reward: 8.30 -> 5.38Hurt
replace reward ensemble with Max-Min policy optimizationEnsemble · RLHFworst-case reward: -1.70 -> -0.10Improved
replace reward ensemble with Max-Min policy optimizationEnsemble · RLHFtrue reward: 2.31 -> 5.38Improved
add linear feature structure (Linear Max-Min vs Max-Min)Max-Min · Trafficlinear worst-case reward: -0.06 -> 0.20Improved
add linear feature structure (Linear Max-Min vs Max-Min)Max-Min · Pandemictrue reward: 1.25 -> 3.65Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy