Mitigating Reward Over-optimization in Direct Alignment Algorithms with Importance Sampling

Re-assign community