Publications

A collection of my research work.

Less Noise, More Voice: Reinforcement Learning for Reasoning via Instruction Purification

Less Noise, More Voice: Reinforcement Learning for Reasoning via Instruction Purification

Yiju Guo, Tianyi Hu, Zexu Sun, Yankai Lin

arXiv:2601.21244 2026

An RLVR framework that boosts sampling success by pruning prompt interference tokens, achieving faster convergence and improved performance.

PDFCode
Learning to Focus: Causal Attention Distillation via Gradient-Guided Token Pruning

Learning to Focus: Causal Attention Distillation via Gradient-Guided Token Pruning

Yiju Guo, Wenkai Yang, Zexu Sun, Ning Ding, Zhiyuan Liu, Yankai Lin

NeurIPS 2025 Conference 2025

A framework to improve LLM reasoning by removing distracting tokens via causal attention distillation and gradient-guided pruning.

PDFCode
LaSeR: Reinforcement Learning with Last-Token Self-Rewarding

LaSeR: Reinforcement Learning with Last-Token Self-Rewarding

Wenkai Yang, Weijie Liu, Ruobing Xie, Yiju Guo, Lulu Wu, Saiyong Yang, Yankai Lin

ICLR 2026 Conference 2025

An efficient RL method that unifies reasoning and verification by utilizing the last-token probability as a self-rewarding signal.

PDFCode
Uncertainty and influence aware reward model refinement for reinforcement learning from human feedback

Uncertainty and influence aware reward model refinement for reinforcement learning from human feedback

Zexu Sun, Yiju Guo, Yankai Lin, Xu Chen, Qi Qi, Xing Tang, Ji-Rong Wen

ICLR 2025 Conference 2025

An uncertainty-aware data augmentation method to refine reward models in RLHF without expensive human annotation.

PDF
Controllable preference optimization: Toward controllable multi-objective alignment

Controllable preference optimization: Toward controllable multi-objective alignment

Yiju Guo, Ganqu Cui, Lifan Yuan, Ning Ding, Zexu Sun, Bowen Sun, Huimin Chen, Ruobing Xie, Jie Zhou, Yankai Lin, others

EMNLP 2024 main conference 2024

A multi-objective alignment method that explicitly controls preference scores to balance helpfulness, honesty, and harmlessness.

PDFCode

Amvae: Asymmetric multimodal variational autoencoder for multi-view representation

Wen Youpeng, Lin Hongxiang, Guo Yiju, Zhao Liang

International Conference on Artificial Neural Networks 2021

A variational autoencoder framework for learning representations from asymmetric multimodal data.