I am a Visiting Researcher at Stanford University, working with Joseph C. Wu and Xuekun Wu on agentic AI for life sciences, and a Student Researcher at Google DeepMind, working with Ming-Hsuan Yang and Liangzhe Yuan on efficient video and vision–language models.
My research connects multi-agent collaboration, reasoning under uncertainty, efficient visual intelligence, and embodied decision-making.
Previously, I worked with teams at Nanyang Technological University, Meta AI, Snap Research, and the HCP Lab of Sun Yat-sen University. Earlier, I worked on risk-control algorithms at Ant Group. More about me ↗
My goal is to help build the next generation of AI agents for scientific discovery.
Expert coordination, learning from failures, and reasoning under uncertainty.
Allocating expert-model assistance across agents by expected benefit and regression risk.
arXiv 2026 (Preprint)
[Paper] [Code]
Treats expert-model assistance across a population of agents as a constrained allocation problem, scoring each agent-service pair by expected gain minus regression risk before any expert output is observed.
Turning recurring failure modes into targeted, verified edits to agent workflows.
ICML 2026 Spotlight
Reframes LLM workflow optimization as reducing a workflow's expected failure mass: CE-Graph clusters counterexamples into failure modes and applies verified graph edits that shrink them.
Matching tasks to expert agents with knowledge-aware Bayesian bandit selection.
KABB selects which expert agents to use for a task via Thompson Sampling guided by a semantic knowledge-distance model, with dual adaptation that keeps expert representations current.
Combining specialized agents and a critic through game-theoretic, uncertainty-aware collaboration.
Base agents handle visual sub-tasks while a critic agent checks consistency; modeled as a non-zero-sum game, they exchange claims and uncertainties and debate when they disagree.
Coherent reasoning, counterfactual learning, and compact visual representations.
Using self-derived coherence and grounding signals to reward multimodal reasoning.
NeurIPS 2026 Spotlight
A label-free reinforcement learning framework that rewards multimodal LLM reasoning chains using five self-derived coherence and grounding signals, improving accuracy and step consistency without a critic.
Compressing visual tokens with continuous features and discrete semantic anchors.
CVPR 2026 Highlight
Compresses a VLM's hundreds of visual tokens into a single token by fusing continuous patch features with a few discrete semantic anchor tokens, retaining most benchmark performance.
Learning from counterfactual image–text pairs to distinguish minimal visual changes.
Fine-tunes vision-language models with counterfactual image-text pairs so they stay aligned, distinguish factual from counterfactual scenes, and respond to minimal causal edits, improving compositional reasoning.
Building a compact video encoder on image representations with selective temporal context.
arXiv 2026 (Preprint)
[Paper] [Code]
Builds a compact video encoder on a frozen image encoder by selecting a fixed token budget across frames and adding temporal context only to selected tokens.
Studying vocabulary scaling and selecting class subsets to improve open-vocabulary prediction.
[Paper] [Code]
CVPR 2026 Highlight
Shows that an open-vocabulary classifier's robustness to new and distractor classes is bounded by its full-vocabulary performance, and fine-tunes on a submodularly chosen class subset.
Action prediction, 3D world models, and cooperation between embodied agents.
Predicting fine-grained robot actions by denoising discrete action tokens.
[Project] [Paper] [arXiv] [Code]
Predicts robot actions by iteratively denoising quantized action tokens with a discrete diffusion process, improving precision and generalization across tasks, scenes, and viewpoints.
Jointly predicting actions and future 3D geometry from a robot foundation model.
arXiv 2026 (Preprint)
Builds a robot world-action model on a 3D foundation backbone that jointly predicts actions and future 3D geometry, improving manipulation success in simulation and on real robots.
Coordinating local robot plans through retrieved cooperation rules, without runtime messaging.
[Paper] [Code]
Coordinates decentralized embodied agents without explicit runtime messaging by retrieving contextual cooperation rules from a Tacit Rule Memory. The rules align local plans with team-level objectives and help agents avoid conflicting actions.
Further work: RaCoT · AAAI 2026 OralDrDiff · EMNLP 2025 OralFlashVLM · ECCV 2026LLM-CASCoMemTrading Visual Uncertainties
An LLM edited 65 token-embedding rows in a pretrained robot policy. Explore the paired rollouts and replay the complete editing session.
“activate the cooking appliances heat source”
Selected held-out example · LIBERO simulation
Lead Organizer · First author of the challenge report
CVPR 2026 · LoViF [Challenge] [Workshop] [Report]
Benchmarking efficient VLMs for multimodal ad quality scoring.
Organizer · First author of the baseline report
ECCV 2026 · LoViF [Challenge] [Workshop] [Report]
Action-conditioned video generation for bimanual robot manipulation.
