SPPO: Sequence-Level PPO for Long-Horizon Reasoning
Reformulates long-horizon reasoning as a sequence-level contextual bandit with a decoupled scalar value function to improve stability and reduce memory cost.
Publications
Research output
Work on reinforcement learning, post-training, and optimization. Use “Copy BibTeX” to copy a paper's citation in one click.
DyJR: Preserving Diversity in Reinforcement Learning with Verifiable Rewards via Dynamic Jensen-Shannon Replay
Introduces Dynamic Jensen-Shannon Replay to preserve diversity in reinforcement learning through verifiable rewards mechanism.
Anchored Policy Optimization: Mitigating Exploration Collapse via Support-Constrained Rectification
Introduces support-constrained rectification for RLVR to mitigate exploration collapse and improve both Pass@1 and response diversity.