I am a master's student in Electronic Information at Minzu University of China and an incoming PhD student at the Data-Centric AI (DCAI) group, Peking University. I am currently looking for research internship opportunities related to LLM post-training, RL for LLMs, and Agent capability construction.

My research focuses on how large language models acquire, improve, and preserve capabilities through post-training. I am interested in reinforcement learning for LLMs, instruction-following robustness, semantic-reward optimization, data and benchmark construction, and Agent capability evaluation. My recent work studies post-training methods such as GRPO-based semantic reward learning, robustness recovery under distracting instructions, automatic benchmark generation for data-analysis agents, and systematic monitoring of Agent potential in pretraining and midtraining models.

Earlier in my research, I worked extensively on multilingual and low-resource language modeling, including datasets, models, and benchmarks for minority languages in China. I now treat these settings as important testbeds for studying capability expansion, alignment preservation, and post-training reliability.

我是中央民族大学电子信息专业硕士生，即将加入北京大学 Data-Centric AI (DCAI) 课题组攻读博士。目前希望寻找大模型后训练、RL4LLM 与 Agent 能力构建相关的科研实习机会。

我的研究关注大语言模型如何通过后训练获得、提升并保持复杂能力，具体包括 RL for LLM、指令遵循鲁棒性、语义奖励优化、数据与 benchmark 构建，以及 Agent 能力评测。近期工作围绕 GRPO 语义奖励强化学习、噪声指令干扰下的鲁棒性恢复、数据分析 Agent benchmark 自动构建，以及 pretrain/midtrain 阶段模型 Agent 潜能监测展开。

早期研究主要集中在多语言与低资源语言建模，包括中国少数民族语言的数据、模型和评测基准建设。现在我更多将这些场景作为研究模型能力扩展、对齐保持与后训练可靠性的实验载体。

• LLM post-training, RLHF/RLVR, GRPO, and semantic-reward optimization
• Instruction following, robustness, and reliability under noisy or conflicting contexts
• Agentic RL, tool-use evaluation, and Agent capability monitoring
• Data synthesis, trajectory construction, and automatic benchmark generation
• Multilingual and low-resource capability expansion as post-training testbeds

• Data-Centric AI (DCAI) Group, Peking University, Research Intern / Research Assistant, 2025.09 - Present
  Working on research and open-source systems related to Data-Centric AI, LLM data systems, Agents, benchmark construction, and AI4Science.

• Foundation Model Group (Ling), Ant Group, Research Intern, 2025.10 - Present
  Working on instruction-following post-training optimization for Ling/Ring foundation models, including failure-mode analysis, training-feedback diagnosis, targeted data construction, and robustness-oriented alignment research.

• Reinforcement Learning with Semantic Rewards Enables Low-Resource Language Expansion without Alignment Tax
  ACL 2026 Findings, first author. A GRPO-based semantic-reward post-training method for capability expansion while preserving general alignment. [arXiv]

• InsightBenchMaker: Towards Generating Evolving and High-Fidelity Benchmarks for Data-Analysis Agents
  ACL 2026 Findings, co-first author. An automatic benchmark construction framework for data-analysis agents with heterogeneous data, executable verification, and high-fidelity insight-data alignment.

• FTibSuite: A Comprehensive Resource Suite for Tibetan Vision-Language Modeling
  ACL 2026 Findings, co-first and corresponding author. A resource suite for Tibetan vision-language modeling, including data, benchmarks, and baseline models. [arXiv]

• Multilingual Encoder Knows more than You Realize: Shared Weights Pretraining for Extremely Low-Resource Languages
  ACL 2025 Main, first author. A shared-weight pretraining framework for adapting multilingual encoders to extremely low-resource generation tasks. [arXiv] [Code/Model]

• CMHG: A Dataset and Benchmark for Headline Generation of Minority Languages in China
  EMNLP 2025 Main, advisor as first author; co-first author. A dataset and benchmark for headline generation in Tibetan, Uyghur, and Mongolian. [arXiv] [Dataset]

• The Curse of Helpfulness: Inverse Scaling Law in Robustness to Distractor Instructions via DistractionIF
  A diagnostic benchmark and analysis of instruction-following robustness under distracting pseudo-instructions in reference text, with GRPO-based robustness recovery. [arXiv]

• Base-model Agent-Potential Monitoring: Probing ReAct-like Interaction Ability in Pretrain and Midtrain Models
  A framework for monitoring Agent-like interaction ability in base and midtraining models through few-shot reasoning, structured action generation, and real environment interaction.

• Source-Grounded Semantic Reinforcement Learning for Low-Resource Target-Language Generation
  A source-grounded semantic RL framework for reference-free low-resource target-language generation. [arXiv]

• Expert Attention: Routing-Guided Static Pruning for Transformer Encoders
  A routing-guided structured pruning method for Transformer encoders.

• Paper2Any: An open-source system that turns papers, images, and text into editable research figures, technical route diagrams, and presentation materials. I contributed to DrawIO/PPT generation, PPTPolish, and knowledge-enhanced generation modules. [GitHub]

• Open-NotebookLM: An open-source document-centered knowledge system supporting semantic retrieval, evidence-grounded QA, AI-assisted note generation, mind maps, slides, podcasts, quizzes, and deep research reports. I contributed to multi-source document ingestion and generation modules for slides, mind maps, DrawIO diagrams, podcasts, quizzes, and deep research reports. [GitHub]

• GitHub: @asd765973346
• Email: rickamorty@muc.edu.cn