Biography

My research interests are in machine learning, optimization, and natural language processing. I have published ~140 papers in ML (NeurIPS, ICML, ICLR), NLP (ACL, EMNLP, NAACL), CV (CVPR, ICCV, ECCV), DM (KDD, ICDM), AI (AAAI, IJCAI) conferences, and journals as Machine Learning (Springer), IEEE TPAMI/TIP/TNNLS/TKDE, etc. Our recent works mainly focus on:

• Human-AI Hybrid Intelligence on Curriculum Learning, Training Dynamics, Unsupervised Exploration, Human-AI Alignment & Teaming, Theory-of-Mind, Co-Education, etc;
• Training Generative AI for better controllability, efficiency, and reasoning skills;
• Synthetic data/tasks, self-evolving creative AI, and auto-benchmarking;
• Neuro-symbolic, Physics-Informed World Models & Embodied Multi-modal Agents;
• Mixture-of-Experts, Multi-Agent, and Collaborative Learning;
• Memorization and Generalization mechanism in Foundation Models.

Our studies are built upon recent LLMs, unified multi-modal models, RL, agentic workflows, to address practical challenges in education, design, medical health, visualization, embodied intelligence, autonomous driving, etc. Our goal is to develop efficient, versatile, trustworthy, and environmentally-friendly hybrid-intelligence based on coevolution between humans and machines. Our code/data/models can be found at Tianyi Lab’s GitHub and HF.

I was a visiting research scientist at Google between 2021-2022, hosted by Prof. Boqing Gong and Prof. Ming-Hsuan Yang. Before that, I received my Ph.D. (thesis) from Computer Science of University of Washington, where I was a member of MELODI lab led by Prof. Jeff A. Bilmes. I have been working with Prof. Dacheng Tao as a research assistant at University of Technology, Sydney (UTS) and Nanyang Technological University. I was a research intern at Yahoo! Labs, mentored by Dr. Hua Ouyang (Apple) and Prof. Yi Chang (Jilin University), and a research intern at Microsoft Research, mentored by Dr. Lin Xiao (Meta AI).

News

• 2025/08: 3 works (Demos’ Position in Prompt bias, Schoenfeld’s Episode Theory for LLM reasoning, NoWait reasoning on LLMs/VLMs) have been accepted by EMNLP 2025.
• 2025/08: I am going to serve as an Area Chair of ICLR 2026 and AAAI 2026.
• 2025/07: 3 works (C3PO, MiP-Overthinking, DynaSaur) have been accepted by COLM 2025.
• 2025/05: We release BLIP3-o, a fully-open sourced framework to unify image reasoning + generation in one model. It aligns the two foundamental tasks via CLIP-level features. You can find report and demo.
• 2025/05: 6 works (3 main with 2 Oral + 3 findings) have been accepted by ACL 2025.
• 2025/05: 3 works have been accepted by ICML: R2-T2 (Test-Time Multimodal MoE), Preference Controllable Multi-Objective RL, and Tilted Sharpness-Aware Minimization.
• 2025/04: We release WALL-E 2.0 (code) to improve the world model for LLM-based embodied agents. TL;DR: WALL-E 2.0 = Neuro-Symbolic World Model + MPC-based LLM Agent, where Neuro-Symbolic World Model = LLM + Complementary Symbolic knowledge (action rules, scene graph, knowledge graph) extracted from LLM errors in experiences.
• 2025/04: We release ColorBench, the first comprehensive benchmark exposing the weaknesses of existing VLMs on color perception, reasoning, and robustness, across >5,800 images from 11 tasks based on real applications including shopping, medical test-kit reading, map-reading, agriculture, art, wildlife research, etc.
• 2025/04: We release C3PO and R2-T2, which develop a novel class of test-time optimization approaches that can significantly boost the performance of existing MoE VLMs and MoE LLMs by 7-15% on challenging tasks.
• 2025/04: We release two works investigating how traning data for LLMs (slow vs. fast thinking, instruction vs. reasoning) lead to different layer-wise gradient patterns and training dynamics, from which we unified vairous existing data quality metrics into one: small but diverse gradient directions indicate better data.
• 2025/04: We release ATLaS and EEF that study better usage of expert trajectories in supervised/reinforcement finetuning of LLM agents. ATLaS selects only 30% critical steps from expert trajectories to achieve better generalization across tasks/envs, while EEF show that expert failed trajectories can considerably improve the agent exploration in RL.
• 2025/04: I am going to serve as an Area Chair of NeurIPS 2025.
• 2025/03: Florence-VL (3B/8B pretrained/SFT VLMs) trained with depth-breadth fusion has been accepted by CVPR 2025 and the complete training recipe has been open-sourced.
• 2025/02: Check the very first “Aha moment” of multimodal reasoning by RL acheived by VisualThinker-R1-Zero from our TurningPoint-AI team!
• 2025/01: 7 ICLR + 3 NAACL accepted, featuring our latest works on synthetic data for post-training, MoE (ICLR Oral), many-objective optimization, in-context transferality, multi-modality imbalance & alignment, oversensitiveness & controllability of GenAI.
• 2024/12: I am going to serve as an Area Chair for ARR Dec 2024 & Feb 2025 (ACL 2025).
• 2024/11: I am going to serve as an Area Chair (SPC) for IJCAI 2025.
• 2024/09: Five papers (3 main + 2 findings) have been accepted by EMNLP 2024.
• 2024/09: I am going to serve as an Area Chair of ICLR 2025.

Research Topics

• Machine Learning (2008-present)
  1. Learning over time: Curriculum Learning, Continual Learning (DisCL, DIH, DoCL, MECE, Time-Consistency, RAR, FPF, CoTASP)
  2. Learning via interactions: Reinforcement Learning, Online Learning (CHER, Unsupervised RL, CO-PILOT, P3)
  3. Learning across tasks/domains: Multi-task Learning, Meta-Learning, Domain Adaptation/Generalization (GPN, PPN, MTC)
  4. Learning multiple models: Mixture-of-Experts (MoE), Collaborative/Cooperative Learning, Federated/Decentralized Learning (MoEE, DivE2, L2C, SCooL, DivFL, FedProto, FedRAP, CAM, MosT)
  5. Learning under noises: Noisy-Label Learning, Adversarial Learning (RoCL, CD-VAE)
  6. Learning representations: Self-Supervised Learning, Dimension Reduction (MEN, LPA3, IDAA)
  7. Sparse Learning: Compressed Sensing, Matrix Factorization, Spectral Method (GoDec, DCA, k-bit HCS)
  8. Optimization: Continuous, Combinatorial, Multi-Objective, Zeroth-order (MosT, Minimax CL, Submodular Partition, TSAM)
• Natural Language Processing (2016-present)
  1. Attention mechanisms: DiSAN, BiBloSA
  2. Data Engineering (selection, exploration, synthesis) for Large language models (LLMs) training: Reflection-Tuning, SuperFiltering, Alpagasus, Cherry LLM, Mosaic-IT, RuleR
  3. LLM Agents, NeuroSymbolic World Models: WALL-E, DynaSaur, ATLaS, EEF
  4. Personalization and Human-AI Alignment: DEBATunE, MCTune, CAIMIRA
  5. Prompting: InstructZero, MoP
  6. In-Context Learning: BenTo, Div-S3
  7. Test-Time Optimization: C3PO, R2-T2
  8. Embedding: MoEE, MetaEOL
  9. Efficient Inference: SpecHub, BumbleBee
  10. Adversarial attack and defense（Jailbreak, Unlearning, etc.): DrAttack
  11. Reasoning: MiP-Overthinking, AdaRFT
• Multi-modal Models (2021-present)
  1. Unified Multi-modal Models with Dense Alignment: Florence-VL, BLIP3-o
  2. Reasoning Multi-modal Models: VisualThinker-R1-Zero
  3. Multi-modal Embodied Agent and RL: EMMA, CoTASP
  4. Multi-modal Generative Agents: MuLan, CoSTA*, FaSTA*
  5. Benchmarks: HallusionBench, AutoHallusion, MOSSBench, ColorBench