Part 1: Why Off-Policy Breaks RL — An SGA Analysis Framework

Part2: Applying the SGA Framework — Token v.s. Sequence-level Correction

Part 3: Trust Region Optimization via Sequence Masking

转载自：

1. https://richardli.xyz/post/rl-collapse-part1/
2. https://richardli.xyz/post/rl-collapse-part2/
3. https://richardli.xyz/post/rl-collapse-part3/