Pillar 01
Research program
Foundation models as decision-making systems.
Furong Lab studies models that allocate computation, choose tools and collaborators, act in physical environments, and revise behavior after failure. Select any pillar or topic to follow the work from research question to project and paper.
Pillar 02
Reasoning control
Decide how to thinkExplore this pillar →Pillar 03
Trustworthy self-improvement
Learn from failureExplore this pillar →Pillar 01
World models
Learn what can happenRepresent physical and digital environments so agents can predict consequences, transfer across embodiments, and plan beyond direct experience.
Pillar 01 flagshipFeatured area
Embodied AI
World models become tangible when agents must understand space, bodies, state, and consequences. Explore representative systems, videos, code, and papers.
Pillar 02
Reasoning control
Decide how to thinkTreat inference as a control problem: allocate computation, collaboration, evidence, and verification where they improve decisions most.
Pillar 02 flagshipFeatured area
AI Agents
Agents turn reasoning into purposeful behavior by planning, using tools, verifying progress, and recovering when an initial strategy fails.
Pillar 03
Trustworthy self-improvement
Learn from failureBuild closed-loop systems that discover failures, intervene before they unfold, and convert uncertainty and hard examples into safer behavior.
Pillar 03 flagshipFeatured area
AI Safety
Adaptive attacks, agentic evaluations, and stress tests expose latent failures so models can be corrected before unsafe behavior reaches deployment.
Research in the world
Methods become meaningful through where they are used.
The lab’s theoretical and systems work has reached problems in science, infrastructure, finance, content integrity, and embodied intelligence.
Biology and medicine
Learning structure in brain-cell populations and human-disease hierarchies, with applications extending to therapeutic discovery.
Robotics and embodied systems
World models and transferable policies that help agents understand scenes, predict interaction, and act through physical bodies.
Resilient infrastructure
Sequential decision-making and multi-agent learning for systems such as power grids operating under uncertainty and disruption.
Financial integrity
Robust, private, and fair learning for financial models exposed to changing data and adversarial behavior.
Content authenticity
Benchmarks, attacks, and defenses for hallucination, data poisoning, AI-generated content, and invisible watermarks.
Efficient adaptation
Methods for updating, routing, and fine-tuning large industrial models while controlling computation and data requirements.
Representative projects
Systems, evidence, and open research artifacts.
Project pages bring together papers, code, demonstrations, and the visual evidence behind the research.
μ0
A scalable world model that predicts semantic 3D interaction traces, learning embodiment-agnostic motion priors from video-only pretraining for downstream robot control.
Agentic Critical Training
An agent-training framework that converts critique and revision into a learning signal for more reliable multi-step behavior.
Research foundations
Earlier work that shaped the current program.
Optimization, latent-structure learning, graphical models, learning theory, fairness, and robustness established ideas that continue through the lab’s current work on foundation models.
Foundational papers
Methods beneath the modern systems.
Selected contributions trace a path from nonconvex optimization and latent-structure recovery to adversarial training, robust sequential decisions, and provable defenses.
Escaping From Saddle Points - Online Stochastic Gradient for Tensor Decomposition
Conference of Learning Theory (COLT) 2015
BibTeX
@inproceedings{ge2015escaping54ac,
title = {Escaping From Saddle Points - Online Stochastic Gradient for Tensor Decomposition},
author = {Rong Ge and Furong Huang and Chi Jin and Yang Yuan. (Alphabetic Order)},
booktitle = {Conference of Learning Theory (COLT) 2015},
year = {2015},
eprint = {1503.02101},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/1503.02101},
}
Are Adversarial Examples Created Equal? A Learnable Weighted Minimax Risk for Robustness under Non-Uniform Attacks
35th AAAI Conference on Artificial Intelligence (AAAI), 2021
BibTeX
@inproceedings{zeng2021adversarialb445,
title = {Are Adversarial Examples Created Equal? A Learnable Weighted Minimax Risk for Robustness under Non-Uniform Attacks},
author = {Huimin Zeng and Chen Zhu and Tom Goldstein and Furong Huang},
booktitle = {35th AAAI Conference on Artificial Intelligence (AAAI), 2021},
year = {2021},
eprint = {2010.12989},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2010.12989},
}
Efficient Adversarial Training without Attacking: Worst-Case-Aware Robust Reinforcement Learning
Advances in Neural Information Processing Systems (NeurIPS), 2022
BibTeX
@inproceedings{liang2022efficientf128,
title = {Efficient Adversarial Training without Attacking: Worst-Case-Aware Robust Reinforcement Learning},
author = {Yongyuan Liang and Yanchao Sun and Ruijie Zheng and Furong Huang},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS), 2022},
year = {2022},
eprint = {2210.05927},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2210.05927},
}
Rethinking Adversarial Policies: A Generalized Attack Formulation and Provable Defense in RL
The Twelfth International Conference on Learning Representations (ICLR), 2024
BibTeX
@inproceedings{liu2024rethinking9da2,
title = {Rethinking Adversarial Policies: A Generalized Attack Formulation and Provable Defense in RL},
author = {Xiangyu Liu and Souradip Chakraborty and Yanchao Sun and Furong Huang},
booktitle = {The Twelfth International Conference on Learning Representations (ICLR), 2024},
year = {2024},
eprint = {2305.17342},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2305.17342},
}
High-Dimensional Gaussian Graphical Model Selection: Walk-Summability and Local Separation Criterion
Journal of Machine Learning Research, Aug. 2012. An abridged version appears in the Conference on Neural Information Processing Systems, Dec. 2011
BibTeX
@article{anandkumar2011high99e0,
title = {High-Dimensional Gaussian Graphical Model Selection: Walk-Summability and Local Separation Criterion},
author = {Animashree Anandkumar and Vincent YF Tan and Furong Huang and Alan S. Willsky},
journal = {Journal of Machine Learning Research, Aug. 2012. An abridged version appears in the Conference on Neural Information Processing Systems, Dec. 2011},
year = {2011},
eprint = {1107.1270},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/1107.1270},
}
High-Dimensional Structure Learning of Ising Models: Local Separation Criterion
Annals of Statistics, Volume 40, Number 3 (2012), 1346-1375. An abridged version appears in the Proc. of NIPS, Dec. 2011
BibTeX
@article{anandkumar2011high711f,
title = {High-Dimensional Structure Learning of Ising Models: Local Separation Criterion},
author = {Animashree Anandkumar and Vincent YF Tan and Furong Huang and Alan S. Willsky},
journal = {Annals of Statistics, Volume 40, Number 3 (2012), 1346-1375. An abridged version appears in the Proc. of NIPS, Dec. 2011},
year = {2011},
eprint = {1107.1736},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/1107.1736},
}
Learning High-Dimensional Mixtures of Graphical Models
Conference on Neural Information Processing Systems (NIPS), 2012
BibTeX
@inproceedings{anandkumar2012learning3bb1,
title = {Learning High-Dimensional Mixtures of Graphical Models},
author = {Animashree Anandkumar and Daniel Hsu and Furong Huang and Sham M. Kakade},
booktitle = {Conference on Neural Information Processing Systems (NIPS), 2012},
year = {2012},
eprint = {1203.0697},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/1203.0697},
}
Learning Deep ResNet Blocks Sequentially using Boosting Theory
The 35th International Conference on Machine Learning (ICML), 2018
BibTeX
@inproceedings{huang2018learning3f73,
title = {Learning Deep ResNet Blocks Sequentially using Boosting Theory},
author = {Furong Huang and Jordan Ash and John Langford and Robert Schapire},
booktitle = {The 35th International Conference on Machine Learning (ICML), 2018},
year = {2018},
eprint = {1706.04964},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/1706.04964},
}