Haowei Lin

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

I am Haowei Lin(林昊苇), a second year Ph.D. student at the Institute for Artificial Intelligence, Peking University, co-advised by Prof. Yitao Liang and Prof. Jianzhu Ma.

I received my Bachelor’s degree in Artificial Intelligence from Yuanpei College, Peking University, where I was fortunate to work with Prof. Bing Liu on OOD detection, continual learning and NLP. We are the first to propose the task of continual pre-training (EMNLP22, ICLR23), and study the theoretical equivalence between OOD detection and continual learning (EMNLP23, ICLR24).

I am passionate about designing next-generation AI that deeply integrates into real world. My primary research focus is in the field of machine learning, with specific interests in Generative Foundation Models (LLM scaling law, 3D autoregressive model, training-free diffusion guidance, discrete flow matching). Currently, I am working on LLM for scientific discovery (e.g., physical law discovery) and complex reasoning (open-world game agent, multi-turn reasoning).

I am a member of Team CraftJarvis, which is dedicated to creating generalist agents for open-world environments. Outside of my professional interests, I enjoy engaging in music-related activities, including singing, playing the guitar, and participating in choirs.

news

May 21, 2025	I’m contributing to the open-source project OpenEvolve, a community implementation of AlphaEvolve—a scientific discovery agent from DeepMind designed to develop better algorithms for open problems. Check out its performance on Symbolic Regression benchmarks!
Dec 01, 2024	Talk on “Unified Training-Free Guidance for Diffusion Models” at NeurIPS 2024 paper sharing session, 机器之心. [video]
Sep 26, 2024	TFG and OmniJARVIS have been accepted at NeurIPS 2024! In TFG, we present a unified training-free guidance method for diffusion models, evaluated across 16 tasks spanning image, audio, and geometry domains. OmniJARVIS, developed in collaboration with Team CraftJARVIS, is an end-to-end VLA (Vision-Language-Action) agent for open-world Minecraft.
May 03, 2024	I will present our new paper Selecting Large Language Model to Fine-tune via Rectified Scaling Law at ICLR 2024 in ME-FoMo workshop. This paper is selected as an oral presentation and is recently accepted by ICML 2024. See you in Vienna!
Jan 16, 2024	Our paper on continual learning has been accepted at ICLR 2024! We propose a theoretically principled and empirically effective method for CL. Feel free to explore our code and paper. This research was conducted during my undergraduate studies under the guidance of Prof. Bing Liu.

selected publications

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.