{"version":1,"id":"han-fang","title":"Han Fang","summary":"Han Fang is an AI research scientist at Meta Superintelligence Labs known for research on agentic capabilities, recursive self-improvement, and machine learning.","content":"**Han Fang** is an artificial intelligence researcher at [Meta Superintelligence Labs](https://iq.wiki/wiki/meta-superintelligence-team), where he works on recursive self-improvement and agent research. He is a core contributor to agent capabilities in Meta’s Muse models, including Muse Spark, and previously founded Meta AI’s production post-training team.[\\[1\\]](#cite-id-b5uwkkkw6d)​[\\[2\\]](#cite-id-a7xsth8xfv)​\n\n## Career\n\n### Meta AI & Muse\n\nFang’s work at Meta includes research on agents that can improve their own capabilities. He describes Muse Spark as a vehicle for research at Meta’s Fundamental AI Research organization and says he contributed to its agentic tool use. \n\nIn July 2026, he reported that Muse Spark achieved state-of-the-art results on MCP-Atlas and frontier performance on Toolathlon. He also described the development of agents that automatically “hill-climb” their capabilities as a memorable result.[\\[1\\]](#cite-id-b5uwkkkw6d)​\n\nFang’s personal website describes self-improving agents that recursively improve models’ agentic capabilities, reporting state-of-the-art performance on MCP-Atlas and frontier performance on Toolathlon and Automation Bench. It also identifies him as a contributor to agent capabilities across Muse Spark, Wonder, Open, and Glimmer.[\\[2\\]](#cite-id-a7xsth8xfv)​\n\nThe site describes Muse Spark Open as an open-weight model with agentic capabilities and says the team shipped Muse, powered by Muse Spark’s agentic capabilities.[\\[2\\]](#cite-id-a7xsth8xfv) \n\nIn October 2026, Fang shared a technical report on Muse Glimmer and described it as a small, nimble, well-rounded model. Meta AI described Muse Glimmer as a 30-billion-parameter open-weight model for agentic tasks on consumer hardware, capable of running at practical speed on a Mac or PC with a single consumer graphics processing unit at 4-bit precision; Meta said it led comparably sized open models in tool use.[\\[3\\]](#cite-id-82h41mcioi)​\n\n### Meta AI & Facebook AI\n\nFrom 2023 to January 2025, Fang was a Senior Staff Research Scientist and technical lead at Meta. He founded Meta AI’s production post-training team and led production post-training for Llama 2 and Llama 3. He said he launched Meta AI in 2023 and helped scale it to one billion monthly active users, working on integrated training runs, core capabilities, tool use, and a data flywheel.[\\[1\\]](#cite-id-b5uwkkkw6d)​[\\[2\\]](#cite-id-a7xsth8xfv)​\n\nFrom October 2019 to January 2023, Fang was a Staff Research Scientist and technical lead at Facebook AI, where he led the chief technology officer’s special projects on multimodal and multilingual content understanding. He developed Meta’s first Few-shot Learner model, which combined policy and content understanding with a reinforcement-learning-based model optimizer. He also developed the Linformer efficient Transformer architecture, which he said analyzed billions of pieces of content per day.[\\[1\\]](#cite-id-b5uwkkkw6d)​\n\nFang was a Senior Research Scientist at Facebook AI from 2017 to 2019. He also interned as a Facebook data scientist in summer 2016.[\\[1\\]](#cite-id-b5uwkkkw6d)​\n\n### Genomics Research\n\nFrom 2014 to 2017, Fang was a research assistant and PhD student in residence at Cold Spring Harbor Laboratory through a joint program with Stony Brook University. His work included developing graphical and machine-learning algorithms for large-scale genomics data. At Cold Spring Harbor Laboratory from 2013 to 2014, he developed algorithms and computational methods for analyzing next-generation sequencing data.[\\[1\\]](#cite-id-b5uwkkkw6d)​\n\nFang’s research has addressed rare diseases, genetic association and epidemiology, genomic variation, chromosomal abnormalities, phylogenetics, and schizophrenia. \n\nHe served on the University of Washington Data Science program’s industry advisory board from 2019 to 2024.[\\[1\\]](#cite-id-b5uwkkkw6d)​\n\n## Research and Publications\n\nFang’s earlier publications include work on statistical methods for genetic association and analysis of sequencing data. With Wei Pan, he coauthored _“A Data-Adaptive Sum Test for Disease Association with Multiple_ [_Common_](https://iq.wiki/wiki/common) _or Rare Variants”_ and _“Powerful multi-marker association tests: unifying genomic distance-based regression and logistic regression,”_ both published in 2010. \n\nLater genomics publications include “Reducing INDEL calling errors in whole genome and exome sequencing data” (2014), “Accurate de novo and transmitted indel detection in exome-capture data using microassembly” (2014), “Indel variant analysis of short-read sequencing data with Scalpel” (2016), “GenomeScope: fast reference-free genome profiling from short reads” (2017), “Scikit-ribo Enables Accurate Estimation and Robust Modeling of Translation Dynamics at Codon Resolution” (2018), and “Accurate detection of complex structural variations using single-molecule sequencing” (2018).[\\[1\\]](#cite-id-b5uwkkkw6d)​\n\nHis AI publications include “Linformer: Self-attention with linear complexity” (2020), “Entailment as few-shot learner” (2021), “The Perfect Blend: Redefining RLHF with Mixture of Judges” (2024), and “Multi-IF: Benchmarking LLMs on Multi-Turn and Multilingual Instructions Following” (2024).[\\[1\\]](#cite-id-b5uwkkkw6d) \n\nHis later listed work includes “Think Smarter not Harder: Adaptive Reasoning with Inference Aware Optimization,” published at ICML in 2025, “Autodata: An Agentic Data Scientist to Create High Quality Synthetic Data,” listed as a 2026 arXiv paper, and “Generalized Parallel Scaling with Interdependent Generations,” listed for ICML 2026.[\\[2\\]](#cite-id-a7xsth8xfv)​\n\nFang’s website also features “The Central Dogma of Artificial Intelligence,” which characterizes intelligence as the compression of experience into generalization; “The RL Environment Field Guide,” which uses Pokémon Red to discuss agent-environment loops, observation spaces, reward design, and credit assignment; and “Post-training 101: A Hitchhiker’s Guide,” which covers supervised fine-tuning, reinforcement learning from human feedback, reward models, and implementation.[\\[2\\]](#cite-id-a7xsth8xfv)​\n\n## Education and Awards\n\nFang earned a PhD in Applied Mathematics and Statistics from Stony Brook University in 2017. \n\nHis dissertation, “Graphical and Machine Learning Algorithms for Large-scale Genomics Data,” received the university’s Woo-Jong Kim Dissertation Award for the department’s best dissertation in Applied Mathematics and Statistics in May 2017. \n\nHe also received the department’s Excellence in Research Award in April 2017 and the President’s Award to Distinguished Doctoral Students, selected from five of the university’s 451 PhD graduates.[\\[1\\]](#cite-id-b5uwkkkw6d)​[\\[2\\]](#cite-id-a7xsth8xfv)​\n\nAfter his doctoral adviser, Mike Schatz, moved to Johns Hopkins University, Fang was a visiting graduate student in computer science there in 2016–2017. \n\nHe earned an MS in Applied Mathematics and Statistics from Stony Brook in 2013 and a BS in Optical Physics from Sun Yat-sen University in 2011. \n\nHis other listed honors include an Outstanding Student Scholarship and a Sun Yat-sen University Scholarship, a Reviewers’ Choice award from the American Society of Human Genetics in 2015, and a University of Washington Summer Institute in Statistics for Big Data scholarship in 2015.[\\[1\\]](#cite-id-b5uwkkkw6d)​\n","categories":[{"id":"people","title":"people"}],"tags":[{"id":"Researcher"},{"id":"Developer"}],"images":[{"id":"QmdzFXjj42HiBHfPVKt7wJmXRbg7ghr6Hz84uN6QrCDe4n","type":"image/jpeg, image/png"}],"media":[],"metadata":[{"id":"references","value":"[{\"id\":\"b5uwkkkw6d\",\"url\":\"https://www.linkedin.com/in/hfang15/\",\"description\":\"Professional profile\",\"timestamp\":1793025643972},{\"id\":\"a7xsth8xfv\",\"url\":\"https://hanfang.info/\",\"description\":\"Personal website\",\"timestamp\":1793025643973},{\"id\":\"82h41mcioi\",\"url\":\"https://x.com/han_fang_\",\"description\":\"Social media profile\",\"timestamp\":1793025643973}]"},{"id":"twitter_profile","value":"https://x.com/han_fang_"},{"id":"linkedin_profile","value":"https://www.linkedin.com/in/hfang15/"},{"id":"website","value":"https://hanfang.info/"},{"id":"github_profile","value":"https://github.com/hanfang"},{"id":"main-image-origin","value":"{\"imageId\":\"QmdzFXjj42HiBHfPVKt7wJmXRbg7ghr6Hz84uN6QrCDe4n\",\"originalId\":\"QmdzFXjj42HiBHfPVKt7wJmXRbg7ghr6Hz84uN6QrCDe4n\"}"},{"id":"commit-message","value":"Create the Han Fang wiki"}],"events":[],"linkedWikis":{"founders":[],"blockchains":[],"speakers":[]},"user":{"id":"0x8af7a19a26d8fbc48defb35aefb15ec8c407f889"},"author":{"id":"0x212cb3f4ae6611054637f9f78f18fb628ad258bb"},"language":"en","operator":{"id":"0x212cb3f4ae6611054637f9f78f18fb628ad258bb"}}