Benhao Huang

Hello! Husky here! I'm an MSML student at CMU, working with the Locus Lab. My research focuses on LLM pre-training, particularly model architectures and training algorithms. I am currently interested in:

Loop Models - weight-tied architectures and iterative dynamics.
Scalable Reasoning - architectures and training methods that improve both reasoning capability and computational efficiency.

Before this, I worked on generative world models with the VCG Group at Harvard and Maitrix at UCSD, LLM-agent-based dataset curation with the TRAIS Lab at UIUC, and AI interpretability with the XAI Lab at SJTU. More details are available in Experience below.

News

2026/04
- Two papers (EqR and FloWM) are accepted to ICML 2026. See you in Seoul!
- Awesome Loop Models is released: a curated list and interactive browser for loop-model papers and technical blogs.
2026/01
The code of Flow Equivariant World Models is released.
2025/11
Tech report of PAN World Model is released.
2025/07
Magica, a Generative World Engine, is now available online.
2025/05
DCA-Bench has been accepted to KDD 2025 DB-Track as an oral paper.

Notes & Thoughts

Short posts, ideas, and quick public notes.

3 posts

Selected Works

See the Research page for the complete list.

Towards Looped Models Done Right — Part II: Rethinking at Fixed Points

Benhao Huang, Chufan Shi, Junlin Chen, Shicheng Wen, Zhengzhong Liu, Eric Xing, Xuezhe Ma

Sep 29th 2026

Institute of Foundation Models

#Loop Models

#Fixed Points

#Efficient Inference

Looped LMs need not pay for every loop: near fixed points, they can scale FLOPs without scaling memory. Fixed-point shortcuts speed up training (pretraining, post-training) and inference (prefill, dec ...

paper

Github

Part I

Towards Looped Models Done Right — Part I: Topology, Input Injection, Recurrent-State Design

Benhao Huang, Chufan Shi, Junlin Chen, Shicheng Wen, Zhengzhong Liu, Eric Xing, Xuezhe Ma

Jul 31st 2026

Institute of Foundation Models Blog

#Loop Models

#LLM Pre-training

#Mixture-of-Experts

Controlled Ouro-to-Huginn ablations isolate loop placement, input injection, and recurrent-state design, then test whether the findings transfer to compute-matched MoE models.

Blog

Part II

Towards Looped Models Done Right — Part I: Topology, Input Injection, Recurrent-State Design

ICML 2026

Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning

Benhao Huang, Zhengyang Geng, Zico Kolter

May 20th 2026

#Scalable Reasoning

#Test-Time Compute

#Equilibrium Models

Equilibrium Reasoners learn task-conditioned attractors, enabling scalable latent reasoning through deeper iterations and broader stochastic trajectories at test time.

paper

Github

Tweet 2

Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning

PAN: A World Model for General, Interactable, and Long-Horizon World Simulation

Jiannan Xiang, Yi Gu, Zihan Liu, Zeyu Feng, Qiyue Gao, Yiyan Hu, Benhao Huang, Guangyi Liu, Yichi Yang, Kun Zhou, Davit Abrahamyan, Arif Ahmad, Ganesh Bannur, Junrong Chen, Kimi Chen, Mingkai Deng, Ruobing Han, Xinqi Huang, Haoqiang Kang, Zheqi Li, Enze Ma, Hector Ren, Yashowardhan Shinde, Rohan Shingre, Ramsundar Tanikella, Kaiming Tao, Dequan Yang, Xinle Yu, Cong Zeng, Binglin Zhou, Hector Liu, Zhiting Hu, Eric P. Xing

Nov 13th 2025

#World Model

#Video Generation

#Long-term Consistency

PAN brings imagination to life — fusing language, action, and vision to simulate the world's evolution with stunning realism and consistency.

technical report

Blog

PAN: A World Model for General, Interactable, and Long-Horizon World Simulation

ICML 2026

Flow Equivariant World Models: Structured Dynamics Outside the Field of View

Hansen Lillemark*, Benhao Huang*, Fangneng Zhan, Yilun Du, T. Anderson Keller

Sep 23rd 2025

NeurReps & SpaVLE Workshops @ NeurIPS 2025

#World Model

#Equivariant Models

#Long-term Consistency

On World Modeling the partially observable dynamics in the environment.

DCA-Bench: A Benchmark for Dataset Curation Agents

Benhao Huang, Yingzhuo Yu, Jin Huang, Xingjian Zhang, Jiaqi W. Ma

May 1st 2025

Data World Workshop @ ICML 2025

#LLM Agent

#Benchmark

#Data-centric AI

A benchmark exploring the performance of LLM Agents on detecting issues in datasets hosted on popular platforms.

Experience & Education

Education

Aug 2025 - Dec 2026 (expected)

M.S. in Machine Learning

Carnegie Mellon University

May 2026 - Aug. 2026

Institute of Foundation Models

Research internship on Loop Models on Scale. Advised by Xuezhe Ma.

Sept. 2025 - Present

Locus Lab, CMU

Pittsburgh, PA, USA

Studying scalable equilibrium models. Advised by Zico Kolter.

Education

Sept 2021 - July 2025

B.E. in Computer Science

Shanghai Jiao Tong University

Mar. 2025 - Aug. 2025

VCG Group, Harvard

Cambridge, MA, USA

Developed Flow Equivariant World Models, studying structured dynamics outside the visible field of view. Advised by Yilun Du, Fangneng Zhan, and Andy Keller.

Jun. 2024 - Mar. 2025

Maitrix.org, UCSD

San Diego, CA, USA

Core contributor to PAN World Models, previously leading the data and game-generation components. Advised by Zhiting Hu.

Mar. 2024 - Jun. 2024

Research Intern @ Moonshot AI

remote

Nov. 2023 - Sept. 2024

TRAIS Lab, UIUC

remote

Developed DCA-Bench, a benchmark for evaluating LLM agents on dataset issue detection and dataset curation. Advised by Jiaqi W. Ma.

Feb. 2023 - Jan. 2024

XAI Lab, SJTU

Shanghai, China

Studied AI interpretability and safety through generalizable interaction primitives in DNNs, using Shapley-value theory. Advised by Quanshi Zhang.

Talks

NICE Workshop on Looped Language Models2026/10/17

Awards & Scholarships

Outstanding Graduate of Shanghai Jiao Tong University2025
National Scholarship2024
Rui Yuan Hong Shan Scholarship, Shanghai Jiao Tong University2023
Shao Qiu Scholarship, Shanghai Jiao Tong University2022

Service

Reviewer: ICML 2026, NeurIPS 2025, KDD 2025