Lei Yang

Semper generalis, semper scholaris.

Lei Yang

Master's Student · Tianjin University · Advisor: Deyi Xiong

My research centers on large language models, especially reinforcement learning for reasoning, long-horizon and code agents, and efficient long-context modeling.

My research aims to make LLMs generalizable and capable: able to explore, recover from cross-domain interference, and solve code-centered tasks over extended trajectories. I currently approach this direction through code-agent data and evaluation, with a longer-term focus on RL algorithms for planning, exploration, and recovery.

Research Interests

RL for LLMs Code Agents Long-Horizon Long-Context LLMs

Contact

Google Scholar GitHub

Research

Selected Directions

Reinforcement Learning for LLM Reasoning

I study how reinforcement learning changes LLM behavior during post-training, especially when exploration collapses or multiple domains interfere with one another.

  • Position-aware use of historical diversity as thinking seeds.
  • Local perturbation theory for cross-domain RL interference.
  • Future work on RL algorithms for long-horizon & code agents.

Code Agents and Code Evaluation

I am interested in long-horizon code agents that understand repositories, reason over tasks, and can be evaluated with high-fidelity feedback instead of only static offline tests.

  • Competitive-programming evaluation through online submission.
  • Data pipelines for code understanding and deployment agents.

Long-Context and Multilingual LLMs

I also work on efficient long-context modeling and scalable continued pre-training, including dense and MoE LLMs for Tibetan language modeling.

  • Divide-and-conquer RoPE scaling factor search for long context.
  • Data curation, continued pre-training, SFT, and preference tuning.
  • Future extension of Tibetan LLMs toward omni-modal models with vision and speech.

Publications

Selected Papers

Lei Yang in red indicates first or co-first authorship. * Equal contribution. † Corresponding author.

A Local Perturbation Theory for Cross-Domain Interference and Recovery in Multi-Domain RL

Lei Yang, Siyu Ding, Deyi Xiong†

A mechanistic study of why single-domain RL updates interfere across domains, using local perturbation analysis to explain selective recovery with short domain refreshes.

Thinking Seeds: Leveraging Historical Diversity for Position-Aware RL in LLMs

Lei Yang, Wei Bi, Chenxi Sun, Renren Jin, Deyi Xiong†

A token-level mix-policy RL framework that reuses historical checkpoints as diverse prefixes while keeping continuations on-policy for stable exploration.

From Curated Data to Scalable Models: Continual Pre-training of Dense and MoE Large Language Models for Tibetan

Lei Yang*, Leiyu Pan*, Bojian Xiong, Renren Jin, Shaowei Zhang, Yue Chen, Ling Shi, Jiang Zhou, Junru Wu, Zhen Wang, Jianxiang Peng, Juesi Xiao, Tianyu Dong, Zhuowen Han, Zhuo Chen, Yuqi Ren, Deyi Xiong†

A full pipeline for Tibetan LLM development, from large-scale curated data to dense and MoE continual pre-training, instruction tuning, and benchmarks.

ERRV: Eliciting Efficient Reasoning through Reasoning Vectors for Policy Optimization in Large Language Models

Zhuowen Han, Lei Yang, Renren Jin, Dan Shi, Chenxi Sun, Deyi Xiong†

Introduces reasoning vectors and targeted policy optimization to shorten reasoning traces while preserving answer accuracy.

DVMap: Fine-Grained Pluralistic Value Alignment via High-Consensus Demographic-Value Mapping

Pengyun Zhu, Yuqi Ren, Zhen Wang, Lei Yang, Deyi Xiong†

Builds demographic-value mappings from high-consensus survey groups to support fine-grained pluralistic alignment and generalization across demographics, countries, and values.

ProBench: Benchmarking Large Language Models in Competitive Programming

Lei Yang, Renren Jin, Ling Shi, Jianxiang Peng, Yue Chen, Deyi Xiong†

A competitive-programming benchmark using real online submissions, problem attributes, and error analyses to evaluate advanced code reasoning.

DCIS: Efficient Length Extrapolation of LLMs via Divide-and-Conquer Scaling Factor Search

Lei Yang, Shaoyang Xu, Jianxiang Peng, Shaolin Zhu, Deyi Xiong†

A divide-and-conquer search method for RoPE scaling factors that extends context length with lower fine-tuning and search cost.

Praetor: A Fine-Grained Generative LLM Evaluator with Instance-Level Customizable Evaluation Criteria

Yongqi Leng, Renren Jin, Yue Chen, Zhuowen Han, Ling Shi, Jianxiang Peng, Lei Yang, Juesi Xiao, Deyi Xiong†

A generative LLM evaluator trained with customizable instance-level criteria for flexible pointwise grading, pairwise comparison, and critiques.

FuxiTranyu: A Multilingual Large Language Model Trained with Balanced Data

Haoran Sun, Renren Jin, Shaoyang Xu, Leiyu Pan, Supryadi, Menglong Cui, Jiangcun Du, Yikun Lei, Lei Yang, Ling Shi, Juesi Xiao, Shaolin Zhu, Deyi Xiong†

An open-source multilingual LLM family trained on balanced data across 43 natural languages and 16 programming languages, with SFT and DPO variants.

Background

Education and Experience

Education

2020.09 - 2024.06

Tiangong University

Bachelor's degree in Network Engineering.

Industry Experience

2026.02 - Present

Baidu · Large Model Algorithm Department

Working on multi-domain reinforcement learning for LLMs, leading to the paper A Local Perturbation Theory for Cross-Domain Interference and Recovery in Multi-Domain RL. Also working on code-agent data construction, filtering, and quality control.

2025.07 - 2026.01

Kuaishou · Foundation LLM and Applications

Worked on reinforcement learning for LLM reasoning, leading to the paper Thinking Seeds: Leveraging Historical Diversity for Position-Aware RL in LLMs.

Honors

Awards & Honors

ICPC Asia Regional Contest, Nanjing Silver Medal
CCPC Mianyang Site Bronze Medal
NOIP Advanced Group Provincial First Prize

CV

Academic CV

My academic CV is available as a PDF.

Open CV