Soft Grpo Soft Thinking Rl Lms

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.

adu2021 Updated 7 repo stars

File contents

adu2021/skillxiv commit 5f4a3ef331

Frequently asked questions

npx get-skill add adu2021/soft-grpo-soft-thinking-rl-lms