W2SPO

New RL method W2SPO improves LLM reasoning with short auxiliary branches

Researchers have developed a new off-policy reinforcement learning method called W2SPO, designed to improve reasoning in large language models. This technique addresses the issue of limited reward contrast in standard m…