Weixin Liang Research @ OpenAI

Mixture-of-Transformers: A Sparse and Scalable Architecture for Multi-Modal Foundation Models

Weixin Liang, Lili Yu, Liang Luo, Srinivasan Iyer, Ning Dong, Chunting Zhou, Gargi Ghosh, Mike Lewis, Wen-tau Yih, Luke Zettlemoyer, Xi Victoria Lin
Transactions on Machine Learning Research (TMLR 2025) 
Paper Twitter Bluesky Code

How can we reduce pretraining costs for multi-modal models without sacrificing quality? At Meta, We introduce Mixture-of-Transformers (MoT), a sparse architecture with modality-aware sparsity for every non-embedding transformer parameter (e.g., feed-forward networks, attention matrices, and layer normalization). MoT achieves dense-level performance with up to 66% fewer FLOPs and was pretrained from scratch at 30B-scale; this TMLR signature paper highlights 2x efficiency gains via modality-aware sparsity.

Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning

Weixin Liang*, Yuhui Zhang*, Yongchan Kwon*, Serena Yeung, James Zou
NeurIPS (2022) 
Paper HTML Poster Website Code

Our new paper explains the intriguing AI ModalityGap: in multi-modal AI, there are large gaps in the representation space separating different data types. We show changing the gap improves zero-shot learning and fairness. Interestingly, modality gaps are created at model initialization and are reinforced by contrastive learning.

Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews

Weixin Liang*, Zachary Izzo*, Yaohui Zhang*, Haley Lepp, Hancheng Cao, Xuandong Zhao, Lingjiao Chen, Haotian Ye, Sheng Liu, Zhi Huang, Daniel A. McFarland, James Y. Zou
International Conference on Machine Learning (ICML 2024)
Oral/top 5% of accepted papers, ICML 2024
Best Presentation Runner-up award at ICSSI 2024 (International Conference on the Science of Science and Innovation), National Academy of Sciences in Washington, DC.  
Paper NY Times Article Twitter Code
Slides PDF Google Slides

Our estimates suggest that 10.6% of ICLR 2024 review sentences and 16.9% for EMNLP have been substantially modified by ChatGPT, with no significant evidence of ChatGPT usage in Nature portfolio reviews. Estimated ChatGPT usage in reviews spikes significantly within 3 days of review deadlines. We present an approach for estimating the fraction of text in a large corpus which is likely to be substantially modified or produced by a large language model (LLM).

Mixture-of-Mamba: Enhancing Multi-Modal State-Space Models with Modality-Aware Sparsity

Weixin Liang, Junhong Shen, Genghan Zhang, Ning Dong, Luke Zettlemoyer, Lili Yu,
SCOPE Workshop @ ICLR (2025)  Oral  
Paper Twitter Bluesky Code

Extension of Mixture-of-Transformers to state-space models (Mamba), successfully pretrained from scratch at 1.4B-scale. In the Transfusion (image+text) setting, Mixture-of-Mamba delivers dense-level performance with just 34.76% of the FLOPs.

LMFusion: Adapting Pretrained Language Models for Multimodal Generation

Weijia Shi, Xiaochuang Han, Chunting Zhou, Weixin Liang, Xi Victoria Lin, Luke Zettlemoyer, Lili Yu,
NeurIPS (2025) 
Paper Twitter

Applied Mixture-of-Transformers in a post-training setting for Llama-3, enabling text+image understanding and generation while preserving language performance.

Adaptive Self-improvement LLM Agentic System for ML Library Development

Genghan Zhang, Weixin Liang, Olivia Hsu, Kunle Olukotun,
International Conference on Machine Learning (ICML 2025) 
Best Paper Award at Deep Learning for Code Workshop, ICLR 2025
Paper Twitter Bluesky

Can LLMs program themselves to run faster? Programming AI accelerators is a major bottleneck in ML. Our self-improving LLM agent learns to write optimized code for new hardware, achieving 3.9x better results.

From Replication to Redesign: Exploring Pairwise Comparisons for LLM-Based Peer Review

Yaohui Zhang, Haijing Zhang, Wenlong Ji, Tianyu Hua, Nick Haber, Hancheng Cao, Weixin Liang
NeurIPS (2025) 
Paper PDF

Explores how pairwise comparisons can improve LLM-based peer review workflows.

ResearchCodeBench: Benchmarking LLMs on Implementing Novel Machine Learning Research Code

Tianyu Hua, Harper Hua, Violet Xiang, Benjamin Klieger, Sang T Truong, Weixin Liang, Fan-Yun Sun, Nick Haber
NeurIPS (2025) 
Paper PDF Project Code Leaderboard

Can today’s LLMs implement tomorrow’s research ideas? We put them to the test. Introducing ResearchCodeBench: 212 tasks from 2024–25 ML papers and code, most released after any model’s training cutoff.

OmicsNavigator: an LLM-driven multi-agent system for autonomous zero-shot biological analysis in spatial omics

Li Yiyao, Nirvi Vakharia, Weixin Liang, Aaron T Mayer, Ruibang Luo, Alexandro E Trevino, Zhenqin Wu
bioRxiv (2025) 
Paper PDF

LLM-driven multi-agent system for zero-shot biological analysis in spatial omics.

Weak-for-Strong: Training Weak Meta-Agent to Harness Strong Executors

Fan Nie, Lan Feng, Haotian Ye, Weixin Liang, Pan Lu, Huaxiu Yao, Alexandre Alahi, James Zou
COLM (2025) 
Paper PDF Code

Trains a weak meta-agent to orchestrate stronger executors for improved performance.

Quantifying large language model usage in scientific papers.

Weixin Liang, Yaohui Zhang, Zhengxuan Wu, Haley Lepp, Wenlong Ji, Xuandong Zhao, Hancheng Cao, Sheng Liu, Siyu He, Zhi Huang, Diyi Yang, Christopher Potts, Christopher D Manning, James Y. Zou
Nature Human Behaviour (2025) 
Best Presentation Runner-up award at ICSSI 2024 (International Conference on the Science of Science and Innovation), National Academy of Sciences in Washington, DC.  
Nature.com PDF Code

Our study finds ~17% of recent CS arXiv papers and ~8% of bioRxiv papers substantially rely on LLM assistance. Higher LLM usage correlates with more frequent preprints, crowded research areas, and shorter papers. Featured in Scientific American, ABC.net.au, Entrepreneur, The Conversation, Tech Xplore, and more.

The Widespread Adoption of Large Language Model-Assisted Writing Across Society

Weixin Liang*, Yaohui Zhang*, Mihai Codreanu, Jiayu Wang, Hancheng Cao, James Y. Zou
Cell Patterns (2025)
Selected for Cell Patterns “Editors’ Picks: Best of 2025”
Paper Code Media

This paper offers the first large-scale measurement of LLM-assisted writing across a diverse range of public online text. We find that AI-generated or AI-edited text is already widespread in domains such as consumer complaints, corporate communications, job postings, and international organization press releases. Our results highlight how LLMs are quietly shaping digital communication and raise urgent questions about transparency, authorship, and the future of human-AI collaboration.

Can large language models provide useful feedback on research papers? A large-scale empirical analysis

Weixin Liang*, Yuhui Zhang*, Hancheng Cao*, Binglu Wang, Daisy Ding, Xinyu Yang, Kailas Vodrahalli, Siyu He, Daniel Smith, Yian Yin, Daniel A. McFarland, James Zou
NEJM AI (2024) 
Paper Twitter Code

With the breakthrough of large language models (LLM) such as GPT-4, there is growing interest in using LLMs to generate scientific feedback on research manuscripts. However, the utility of LLM-generated feedback has not been systematically studied. To address this gap, we created an automated pipeline using GPT-4 to provide comments on the full PDFs of scientific papers. Our results suggest that LLM and human feedback can complement each other. While human expert review is and should continue to be the foundation of rigorous scientific process, LLM feedback could benefit researchers, especially when timely expert feedback is not available and in earlier stages of manuscript preparation before peer-review.

Navigating Dataset Documentation in AI: A Large-Scale Analysis of Dataset Cards on Hugging Face

Xinyu Yang*, Weixin Liang*, James Zou
ICLR 2024 (2024) 
PDF Code

Advances in machine learning are closely tied to the creation of datasets. While data documentation is widely recognized as essential to the reliability, reproducibility, and transparency of ML, we lack a systematic empirical understanding of current dataset documentation practices. By analyzing all 7,433 dataset documentation on Hugging Face, our investigation provides an overview of the Hugging Face dataset ecosystem and insights into dataset documentation practices.

Systematic analysis of 32,111 AI model cards characterizes documentation practice in AI.

Weixin Liang, Nazneen Rajani, Xinyu Yang, Ezinwanne Ozoani, Eric Wu, Yiqun Chen, Daniel Scott Smith, James Zou
Nature Machine Intelligence (2024) 
Paper

The rapid proliferation of AI models has underscored the importance of thorough documentation, as it enables users to understand, trust, and effectively utilize these models in various applications. Although developers are encouraged to produce model cards, it's not clear how much information or what information these cards contain. In this study, we conduct a comprehensive analysis of 32,111 AI model documentations on Hugging Face, a leading platform for distributing and deploying AI models. Our investigation sheds light on the prevailing model card documentation practices. Most of the AI models with substantial downloads provide model cards, though the cards have uneven informativeness. We find that sections addressing environmental impact, limitations, and evaluation exhibit the lowest filled-out rates, while the training section is the most consistently filled-out. Our study opens up a new perspective for analyzing community norms and practices for model documentation through large-scale data science and linguistics analysis.

GPT detectors are biased against non-native English writers

Weixin Liang*, Mert Yuksekgonul*, Yining Mao* Eric Wu*, James Zou
Cell Patterns (2023) 
Paper Cell.com Code Data

We should be very cautious when using detectors to classify if text is written by AI or human. Our research has shown that such detectors classify over 50% of real text written by non-native English speakers as AI-generated, while most polished essays generated by GPT evade detection. This creates a bias and false positives against non-native speakers, as literary language is often classified as "human."

Media Coverage: The Markup, The Guardian, Fortune, Stanford HAI, Stanford EE, The Hechinger Report

Accuracy on the Curve: On the nonlinear correlation of ML performance between data subpopulations

Weixin Liang*, Yining Mao* Yongchan Kwon*, Xinyu Yang, James Zou
International Conference on Machine Learning (ICML 2023) 
Paper Poster Website Recording Code

Recent works empirically find that there is a strong linear relationship between in-distribution (ID) and out-of-distribution (OOD) performance, but we show that this is not necessarily true if there are subpopulation shifts. In this paper, we empirically show that out-of-distribution performance often has nonlinear ("moon shape") correlation with in-distribution performance under subpopulation shifts.

OpenDataVal: a Unified Benchmark for Data Valuation

Kevin Jiang*, Weixin Liang*, James Zou, Yongchan Kwon
NeurIPS Datasets and Benchmarks Track (2023) 
Paper Project Page Code Twitter

Assessing the quality and impact of individual data points is critical for improving model performance and mitigating undesirable biases within the training dataset. In this paper, we introduce OpenDataVal, an easy-to-use and unified benchmark framework that empowers researchers and practitioners to apply and compare various data valuation algorithms.

Characterizing the Clinical Adoption of Medical AI Devices through U.S. Insurance Claims

Kevin Wu, Eric Wu, Brandon Theodorou, Weixin Liang, Christina Mack, Lucas Glass, Jimeng Sun, James Zou
NEJM AI (2023) 
Paper

We analyze billions of insurance claims data to produce a first-look at medical AI adoption.

Media Coverage: Nature Medicine, The Imaging Wire, Cardiac Wire

Advances, opportunities and challenges in creating data for trustworthy AI

Weixin Liang, Girmaw Abebe Tadesse, Daniel Ho, Li Fei-Fei, Matei Zaharia, Ce Zhang, James Zou
Nature Machine Intelligence (2022) 
Paper Nature.com Twitter

As AI model-building becomes more automated, much of the resources and time in practice are devoted to designing what data to collect, data cleaning, annotations and data evaluations. Our article discusses the best practices, new challenges and opportunities for each of these key components of the data for AI pipeline.

Disparities in Dermatology AI Performance on a Diverse, Curated Clinical Image Set

R Daneshjou*, K Vodrahalli*, W Liang*, R Novoa, M Jenkins, V Rotemberg, J Ko, S Swetter, E Bailey, O Gevaert, P Mukherjee, M Phung, K Yekrang, B Fong, R Sahasrabudhe, Albert Chiou, James Zou
Science Advances (2022) 
Machine Learning for Health (ML4H 2021) 
Paper Science.org Diverse Dermatology Images (DDI) Dataset
Training physicians and algorithms in dermatology diversity | Scope

In order to train and test AI algorithms in dermatology, we need diverse, validated benchmarks. We curated the Diverse Dermatology Images (DDI) dataset to meet this need—the first publicly available, expertly curated, and pathologically confirmed image dataset with diverse skin tones.

Systematic analysis of 50 years of Stanford University Technology Transfer and Commercialization

Weixin Liang, Scott Elrod, Daniel A. McFarland, James Zou
Patterns (2022) 
Paper Cell.com Twitter Recording
Stanford HAI News: Analyzing 50 Years of Stanford Patents
Finding patterns of success across 50 years of innovation | Scope
OTL 50th Anniversary Report: A Half Century of Pioneering Innovation

Computational analysis of 4,512 inventions marketed by Stanford's Office of Technology Licensing between 1970 and 2020 characterizes how the academic innovation landscape changed over time. We identified factors, such as the composition of the inventors, associated with the commercial success of the inventions. We also identified linguistic differences in how high-revenue and low-revenue inventions in the same field are described and marketed.

MetaShift: A Dataset of Datasets for Evaluating Contextual Distribution Shifts and Training Conflicts

Weixin Liang, James Zou
Contributed Talk at ICML 2022 Workshop on Shift happens: Crowdsourcing metrics and test datasets beyond ImageNet
International Conference on Learning Representations (ICLR 2022) 
Paper HTML Website Code
HuggingFace Recording Blog

MetaShift introduces a collection of >10K sets of images with annotated contexts! Context is missing in many ML datasets but is critical for understanding model performance. It enables evaluating how ML works in different contexts (e.g. indoor cat vs outdoor cat).
Bonus: we give distance between contexts.

SEAL: Interactive Tool for Systematic Error Analysis and Labeling

Nazneen Rajani, Weixin Liang, Lingjiao Chen, Meg Mitchell, James Zou
Empirical Methods in Natural Language Processing (EMNLP 2022) 
Paper SEAL toolkit and Demo

Machine learning systems that seemingly perform well on average can still make systematic errors on important subsets of data. We introduce an interactive Systematic Error Analysis and Labeling (SEAL) tool that uses a two-step approach to first identify high error slices of data and then in the second step introduce methods to give human-understandable semantics to those under-performing slices.

Improving Out-of-Distribution Robustness via Selective Augmentation

Huaxiu Yao, Yu Wang, Sai Li, Linjun Zhang, Weixin Liang, James Zou, Chelsea Finn
International Conference on Machine Learning (ICML 2022) 
Paper HTML Poster Code Recording

To deploy machine learning algorithms in real-world applications, we must pay attention to distribution shift, i.e. when the test distribution is different from the training distribution, which substantially degrades model performance. We propose a simple mixup-based method to learn invariant functions via selective augmentation.

Neural Group Testing to Accelerate Deep Learning

Weixin Liang, James Zou
International Symposium on Information Theory (ISIT 2021) 
Paper HTML Code Slides

Our new ISIT 2021 paper proposes neural group testing to speed up DeepLearning. The idea is to adaptively apply the network to groups of data pooled at suitable layers, which greatly reduces total compute.

HERALD: An Annotation Efficient Method to Train User Engagement Predictors in Dialogs

Weixin Liang*, Kaihui Liang*, Zhou Yu
Annual Conference of the Association of Computational Linguistics (ACL 2021)  Paper HTML Code Recording

We propose a workflow that automatically labels training data with minimum human efforts involved, built upon our previous ACL 2020 work.

ALICE: Active Learning with Contrastive Natural Language Explanations

Weixin Liang, James Zou, Zhou Yu
Empirical Methods in Natural Language Processing (EMNLP 2020) 
Paper HTML Slides Recording Blog

Review Ratings: 4, 4, 4.5 in 5-point scale

Our new EMNLP paper shows how to teach ML via natural language explanation of contrasts between concepts (e.g. "difference between COVID and flu is ..."). It's much more efficient than using labeled examples. Excited for more human-like learning!

Beyond User Self-Reported Likert Scale Ratings: A Comparison Model for Automatic Dialog Evaluation

Weixin Liang, James Zou, Zhou Yu
Annual Conference of the Association for Computational Linguistics (ACL 2020) 
Paper HTML Code Slides Recording Blog

Review Ratings: 4.5, 4.5, 5 in 5-point scale

For dialog system evaluation, we found that self-reported dialog ratings are skewed, noisy and insensitive due to bias and variance among different users. We propose a three-stage denoising pipeline to reduce self-reported ratings and, at the same time, build an automatic comparison-based automatic dialog quality predictor.

MOSS: Training End-to-End Dialog Systems with Modular Supervision

Weixin Liang*, Youzhi Tian*, Chengcai Chen, Zhou Yu
AAAI Conference on Artificial Intelligence (AAAI 2020) 
Paper HTML Slides Press

We propose an end-to-end framework for task-oriented dialog systems, which can flexibly incorporate supervision from multiple intermediate dialog system modules (e.g. natural language understanding, dialog state tracking, dialog policy learning and natural language generation) in an end-to-end manner.

DeepStore: In-Storage Acceleration for Intelligent Queries

VS Mailthody, Z Qureshi, W Liang, Z Feng, SG De Gonzalo, Y Li, H Franke, J Xiong, J Huang, Wen-mei Hwu
International Symposium on Microarchitecture (MICRO 2019) 
Paper

A computer architecture conference paper on in-storage hardware acceleration for deep learning.

CU-Net: Component Unmixing Network for Textile Fiber Identification.

Zunlei Feng, Weixin Liang, Daocheng Tao, Li Sun, Anxiang Zeng, Mingli Song
International Journal of Computer Vision (IJCV 2019) 
Paper

We are the first to leverage computer vision techinques for image-based nondestructive textile fiber identification, which is practically useful in fashion, decoration, and design industry. Existing methods based on physical, chemical and microscopy techniques are normally limited by their long identification cycles, many human factors, high technological barriers, and existing damage.

MemCloak: Practical Access Obfuscation for Untrusted Memory

Weixin Liang, Kai Bu, Ke Li, Jinhong Li, Arya Tavakoli
Annual Computer Security Applications Conference (ACSAC 2018) 
Paper

Outstanding Graduation Thesis, Zhejiang University

Access patterns over untrusted memory have long been exploited to infer sensitive information like program types or even secret keys. We propose a light-weight obfuscation solutions to hide real memory accesses.