9 papers
Geometric Self-Distillation for Reasoning Generalization
Josip JukiÄ, Ivan Titov
On-policy distillation is a practical post-training recipe for large language models, supplying dense teacher supervision on the student's own trajectories. In privileged-context s…
Blending Supervised and Reinforcement Fine-Tuning with Prefix Sampling
Zeyu Huang, Tianhao Cheng, Zihan Qiu +4
Existing LLMs-post-training techniques are broadly categorized into supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT). Each paradigm presents a distinct trade-off: (…
The Cancellation Hypothesis in Critic-Free RL: From Outcome Rewards to Token Credits
Tianhao Cheng, Zeyu Huang, Zihan Qiu +5
A commonly accepted explanation of critic-free RL for LLMs, based on sequence-level rewards, is that it reinforces successful rollouts with a positive advantage while penalizing fa…
A Unified View of Attention and Residual Sinks: Outlier-Driven Rescaling is Essential for Transformer Training
Zihan Qiu, Zeyu Huang, Kaiyue Wen +16
We investigate the functional role of emergent outliers in large language models, specifically attention sinks (a few tokens that consistently receive large attention logits) and r…
A Controllable Examination for Long-Context Language Models
Yijun Yang, Zeyu Huang, Wenhao Zhu +4
Existing frameworks for evaluating long-context language models (LCLM) can be broadly categorized into real-world applications (e.g, document summarization) and synthetic tasks (e.…
Post-hoc Reward Calibration: A Case Study on Length Bias
Zeyu Huang, Zihan Qiu, Zili Wang +2
Reinforcement Learning from Human Feedback aligns the outputs of Large Language Models with human values and preferences. Central to this process is the reward model (RM), which tr…