SofT-GRPO: Surpassing Discrete-Token LLM Reinforcement Learning via Gumbel-Reparameterized Soft-Thinking Policy Optimization
Paper • 2511.06411 • Published • 18 •
2
Zhi Zheng
zz1358m
AI & ML interests
LLM reasoning, Trustworthy LLM, LLM application, Neural combinatorial optimization.
Recent Activity
upvoted a paper 13 days ago
Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents upvoted a paper 15 days ago
Inference-Time Attribute Distribution Alignment for Unconditional DiffusionOrganizations
commented a
paper 8 months ago