soft-grpo-soft-thinking-rl-lms

【效率与协作】Enable policy gradient optimization on soft LLM tokens by injecting Gumbel noise and applying Gumbel-Softmax reparameterization—allowing soft-thinking patterns to match discrete-token RL performance while maintaining continuous optimization advantages.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准