DiDPO: Diff-in-Diff Policy Optimization for Coding Agent Training
The episode discusses DiDPO, a method for training coding agents with reinforcement learning. It addresses credit assignment by splitting code diffs into sub-diffs, grouping similar ones across rollouts, and assigning local advantages. DiDPO outperforms baselines on benchmarks like APPS and USACO, with minimal overhead, and is open-sourced.