Haowei Lin

E-mail: linhaowei (at) pku (dot) edu (dot) cn

I am Haowei Lin, a third year Ph.D. student at the Institute for Artificial Intelligence, Peking University, co-advised by Prof. Yitao Liang, Prof. Di He, and Prof. Jianzhu Ma. I am currently interning with the Agent team at Tencent Hunyuan.

I received my Bachelor’s degree in Artificial Intelligence from Yuanpei College, Peking University. During my undergraduate studies, I was fortunate to work with Prof. Bing Liu and Dr. Zixuan Ke on OOD detection, continual learning, and language models. We are the first to propose the task of continual pre-training for LLMs (EMNLP22, ICLR23), and the first to apply OOD detection methods to continual learning (EMNLP23, ICLR24).

My Ph.D. research focuses on AI for Science, spanning two complementary directions. On one hand, I study generative foundation models across modalities such as text, video, MDPs, 3D, and molecules. In collaboration with Prof. Stefano Ermon and Prof. Yilun Du, we conducted a series of studies on inference-time scaling for non-autoregressive models (the TFG series: NeurIPS24, ICLR25, ICLR26). On the other hand, I am interested in building AI Scientists and treating AI research itself as a new axis for AI4Science. As an initial step, we studied Scaling Laws Discovery (ICML24, ICLR26) to unveil the physics of AI.

Beyond research, I enjoy contributing to open source. I co-lead Harbor Adapters, a subproject of Terminal Bench, an evaluation harness widely used in modern model releases (e.g., GPT, Gemini, Claude).

If you’re interested in collaborating on generative foundation models or AI Scientist research, feel free to reach out via email.

news

Jul 10, 2026	Excited to share Hyra (Hunyuan Research Agent), an autonomous research agent I lead at Tencent Hunyuan. Hyra sets new state-of-the-art results on AI4AI, AI4Science, mathematics, engineering, and creative design (see results).
Jul 07, 2026	Two new benchmarks in the Terminal-Bench series are out: Harbor-Index, a challenging, diverse, and high-quality benchmark for evaluating frontier agents, which I co-lead; and Frontier-Bench (formerly Terminal-Bench 3.0), for which I served as a task contributor and reviewer.
Dec 07, 2025	Glad to launch a new blog on Scaling Law Discovery (SLD) (paper). We hope our work on SLD helps advance foundation model development and push the boundaries of AI Scientist. Code, dataset, benchmarks, and leaderboard are all publicly available.

selected publications

ICML

UniCode: Augmenting Evaluation for Code Reasoning

Xinyue Zheng^*, Haowei Lin^*, Shaofei Cai, and 3 more authors

In The Forty-third International Conference on Machine Learning (ICML 2026),

Abs PDF

UniCode is a generative coding-evaluation framework that creates augmented problem variants, automated tests, and fine-grained error metrics to expose LLM reasoning and scalability weaknesses.
ICLR

Terminal-bench: Benchmarking agents on hard, realistic tasks in command line interfaces

Mike A. Merrill^*, Alexander G. Shaw^*, Nicholas Carlini, and 82 more authors

The Fourteenth International Conference on Learning Representations (ICLR 2026),

Abs PDF Code Website

A carefully curated hard benchmark composed of 89 tasks in computer terminal environments inspired by problems from real workflows. This benchmark is used by OpenAI, GDM, Anthropic model card.
ICLR

Can Language Models Discover Scaling Laws?

Haowei Lin^*, Haotian Ye^*, Wenzheng Feng, and 8 more authors

The Fourteenth International Conference on Learning Representations (ICLR 2026),

Abs PDF Code Website

Automatic AI research. Using code agents to discover scaling law.
Nat. Mach. Intell.

A Neural Symbolic Model for Space Physics

Jie Ying^*, Haowei Lin^*, Chao Yue^*, and 7 more authors

Nature Machine Intelligence (Cover Paper),

Abs PDF Code

Using large-scale synthetic data to train autoregressive models for symbolic regression in space physics.
ICML Spotlight

MCU: An Evaluation Framework for Open-Ended Game Agents

Xinyue Zheng^*, Haowei Lin^*, Kaichen He, and 5 more authors

In The Forty-second International Conference on Machine Learning (ICML 2025),

Abs PDF Code

The first evaluation framework for open-ended game agents, with a task generator to synthesize infinite open-ended tasks and VLM-based evaluator.
ICLR

TFG-Flow: Training-free Guidance in Multimodal Generative Flow

Haowei Lin^*, Shanda Li^*, Haotian Ye, and 4 more authors

In The Thirteenth International Conference on Learning Representations (ICLR 2025),

Abs PDF Code

Adapting training-free guidance method from continuous diffusion to multimodal flow matching (discrete + continuous).
NeurIPS Spotlight

TFG: Unified Training-Free Guidance for Diffusion Models

Haotian Ye^*, Haowei Lin^*, Jiaqi Han^*, and 6 more authors

In Advances in Neural Information Processing Systems 37 (NeurIPS 2024),

Abs PDF Code

The first unified framework for training-free guidance in diffusion models and comprehensive benchmarks on image, audio, molecules.
ICML

Selecting Large Language Model to Fine-tune via Rectified Scaling Law

Haowei Lin^*, Baizhou Huang^*, Haotian Ye^*, and 7 more authors

In The Forty-first International Conference on Machine Learning (ICML 2024),

Abs PDF Website

The first paper to introduce the problem of LLM selection. Proposed to use scaling law for LLM selection. Rectified vanilla power law in LLM fine-tuning scaling.
ICLR

Class Incremental Learning via Likelihood Ratio-Based Task Prediction

Haowei Lin, Yijia Shao, Weinan Qian, and 3 more authors

In The Twelfth International Conference on Learning Representations (ICLR 2024),

Abs PDF Code

A theory-guided class incremental learning method, being state-of-the-art for more than 3 years (almost reaching performance upper bound).
ICLR

Continual Pre-Training of Language Models

Zixuan Ke^*, Yijia Shao^*, Haowei Lin^*, and 3 more authors

In The Eleventh International Conference on Learning Representations (ICLR 2023),

Abs PDF Code Poster

The first paper to introduce the concept of continual pre-training (CPT), with comprehensive benchmarks and a new method motivated by model sparsity.