today local_bar

Songjun Tu

PhD student, Institute of Automation Chinese Academy of Sciences

2 papers at NeurIPS 2025

OpenReview· Semantic Scholar· Google Scholar

Poster Session 5

Friday, December 5, 2025 · 11:00 AM → 2:00 PM

Exhibit Hall C,D,E

AlphaDecay: Module-wise Weight Decay for Heavy-Tailed Balancing in LLMs

#3405 · Di He, Songjun Tu, Ajay Jaiswal, Li Shen, Ganzhao Yuan, Shiwei Liu, Lu Yin

We propose AlphaDecay, a per-module weight decay method guided by heavy-tailedness, improving large language model performance.

Poster Session 6

Friday, December 5, 2025 · 4:30 PM → 7:30 PM

Exhibit Hall C,D,E

Learning When to Think: Shaping Adaptive Reasoning in R1-Style Models via Multi-Stage RL

#5003 · Songjun Tu, Jiahao Lin, Qichao Zhang, Xiangyu Tian, Linjing Li, Xiangyuan Lan, Dongbin Zhao