Tag
Introduces ACPO, a token-level credit assignment framework for reinforcement learning in LLMs that uses fine-grained surrogate entropy to improve reasoning performance on math and coding benchmarks, outperforming strong baselines like DAPO, GTPO, and SAPO.