tracked author
Jiashun Liu (刘嘉顺)
香港科技大学 Ling Pan 研究组博士生与 Tricks or Traps 共同贡献作者，研究覆盖深度强化学习可塑性、GFlowNets 和大语言模型强化学习后训练；早期公开论文记录包含 Tianjin University 机构关系。
Representative Papers
来自作者已核验个人主页的重点论文；本站单篇归档见下方 Related Notes。
- 01 Tricks or Traps? A Deep Dive into RL for LLM Reasoning ICLR · 2026
- 02 Asymmetric Proximal Policy Optimization: mini-critics boost LLM reasoning ICLR · 2026
- 03 Let It Flow: Agentic Crafting on Rock and Roll arXiv · 2025
- 04 Part II: ROLL Flash -- Accelerating RLVR and Agentic Training with Asynchrony arXiv · 2025
- 05 Measure gradients, not activations! Enhancing neuronal activity in deep reinforcement learning NeurIPS · 2025
- 06 The Courage to Stop: Overcoming Sunk Cost Fallacy in Deep Reinforcement Learning ICML · 2025
- 07 Neuroplastic Expansion in Deep Reinforcement Learning ICLR · 2025
- 08 Flow Factorization for Efficient Generative Flow Networks AAAI · 2025
- 09 Unlock the Intermittent Control Ability of Model Free Reinforcement Learning NeurIPS · 2024
- 10 Unlock the Cognitive Generalization of Deep Reinforcement Learning via Granular Ball Representation ICML · 2024
- 11 Hybrid CtrlFormer: Learning Adaptive Search Space Partition for Hybrid Action Control via Transformer-based Monte Carlo Tree Search UAI · 2024
Related Notes
按论文归档时间排序，展示该作者在本站已经出现的材料。
归档 更新
Part I: Tricks or Traps? A Deep Dive into RL for LLM Reasoning
在同一训练框架下、覆盖范围并不完全重合的 Qwen3 4B/8B Base 与对齐后数学强化学习实验中分别检查优势归一化、概率比裁剪、损失聚合和超长过滤，并将组内均值—批级标准差归一化与 token 级损失组成不训练价值模型的 Lite PPO；OpenReview 补充的一个 Qwen3 8B Base 三随机种子设置仍优于 GRPO 与 DAPO，但奖励范围异常、计算量未对齐和有限模型任务覆盖限制普适结论。