Yaorui Shi

I am a 3rd year PhD student at the University of Science and Technology of China (USTC), under the supervision of Prof. An Zhang and Xiang Wang.

My research interests include large reasoning models, reinforcement learning, agentic memory, and agentic self-evolving. I am particularly interested in building agents that can remember, reason, and improve themselves over long horizons.

Before joining USTC, I received my Bachelor’s degree in Computer Science from XJTU’s Young Gifted Class (少年班). Currently, I am a research intern at Meituan LongCat Team, where I lead projects about multi-modal agentic RL.

I am open to research collaborations and internship opportunities, especially those related to agentic memory systems and agentic self-evolving.

news

May 07, 2026	Latest research Skill1 released — unified evolution of skill-augmented agents via RL. Check it out on arXiv and Huggingface 🤗.
May 01, 2026	MemOCR accepted to ICML 2026 🎉🎉🎉 See you at Seoul!
Dec 26, 2025	Latest research paper MemOCR at Meituan released on arXiv and Huggingface 🤗.
Dec 26, 2025	One paper about agentic memory accepted to ICLR 2026 🎉🎉🎉 See you in Brazil!
Sep 23, 2025	One paper about deep search agent accepted to NeurIPS 2025 🎉🎉🎉

selected publications

Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning

Yaorui Shi, Yuxin Chen, Zhengxi Lu, and 6 more authors

arXiv preprint arXiv:2605.06130, 2026

Abs arXiv PDF Code 🤗 HF

A persistent skill library allows language model agents to reuse successful strategies across tasks. We propose Skill1, a framework that trains a single policy to co-evolve skill selection, utilization, and distillation toward a shared task-outcome objective via reinforcement learning. Experiments on ALFWorld and WebShop show that Skill1 outperforms prior skill-based and RL baselines.
MemOCR: Layout-Aware Visual Memory for Efficient Long-Horizon Reasoning

Yaorui Shi, Shugui Liu, Yu Yang, and 7 more authors

arXiv preprint arXiv:2601.21468, 2026

Abs arXiv PDF Code Website 🤗 HF

MemOCR is a multimodal memory agent that enhances long-horizon reasoning by adaptively compressing interaction histories into visual layouts, enabling efficient context utilization under tight budget constraints.
LongCat-Flash-Thinking-2601 Technical Report

Meituan LongCat Team, Anchun Gui, Bei Li, and 8 more authors

arXiv preprint arXiv:2601.16725, 2026

arXiv PDF
Look Back to Reason Forward: Revisitable Memory for Long-Context LLM Agents

Yaorui Shi, Yuxin Chen, Siyuan Wang, and 5 more authors

In International Conference on Learning Representations (ICLR), 2025

Abs arXiv PDF Code

Large language models struggle with long-context question answering when relevant evidence is dispersed across millions of tokens. We propose ReMemR1, which incorporates memory retrieval capabilities into the memory update process, enabling agents to selectively revisit historical memories for non-linear reasoning.
Search and Refine During Think: Facilitating Knowledge Refinement for Improved Retrieval-Augmented Reasoning

Yaorui Shi, Sihang Li, Chang Wu, and 5 more authors

In Advances in Neural Information Processing Systems (NeurIPS), 2025

Abs arXiv PDF Code Website

We introduce AutoRefine, a novel framework that enhances retrieval-augmented reasoning by integrating autonomous knowledge refinement into the reasoning process. AutoRefine dynamically refines retrieved documents to improve retrieval quality and reduce hallucination, enabling more reliable and contextually relevant reasoning.
SciLitLLM: How to Adapt LLMs for Scientific Literature Understanding

Sihang Li, Jin Huang, Jiaxi Zhuang, and 7 more authors

In International Conference on Learning Representations (ICLR), 2025

Abs arXiv PDF Code Website

We address challenges in using large language models for scientific document comprehension by proposing a hybrid strategy that integrates continual pre-training with supervised fine-tuning to infuse scientific knowledge while improving task-specific instruction-following abilities.