Foundation model training
Large-scale pre-training and post-training—including SFT and reinforcement learning—for conversational and multimodal foundation models.
I work across the full stack of foundation model capability.
pre-training to build representations, SFT and RL to shape behavior, synthetic data and evaluation to expose failure modes, and agents and tools to turn intelligence into reliable action.
Currently, I work on conversational and multimodal foundation models; previously, I built multimodal AI agents for content creation on Apple Vision Pro. I studied computer systems at Carnegie Mellon University.
My mission is to turn foundation models from fluent predictors into reliable collaborators — systems that learn deeply, act intelligently, and interact naturally with people.
I write about how training shapes model behavior, how agents learn to use tools, and how AI can expand human thought without replacing human judgment. I am also a reviewer for NeurIPS and ACL.
Open to research conversations around foundation model training, post-training, agentic systems, and human-AI interaction. Reach me by email, connect on LinkedIn, or explore my work on Google Scholar.
Current Focus
From fluent predictors to reliable collaborators. I study how models acquire capability, how training shapes their behavior, and how intelligent systems can interact with people naturally.
Data, systems & evaluation
Synthetic-data pipelines, simulated users, test-time scaling, and evaluations on distributed infrastructure.
Agents & interaction
Agents, tool capability memory, guardrails, and full-duplex interaction that turn model intelligence into dependable human–AI collaboration.
Research Publications
01ToolMem: Enhancing Multimodal Agents with Learnable Tool Capability MemoryA dynamic memory for agents to learn the strengths and weaknesses of neural tools from test time experience. ToolMem improves tool-performance prediction by 14.8% in text settings and 28.7% in multimodal settings, and improves optimal tool selection by 21 and 24 percentage points.02MMPerspective: Do MLLMs Understand Perspective?A 10-task benchmark with 2,711 images and 5,083 question-answer pairs for testing perspective perception, spatial reasoning, and robustness across 43 multimodal large language models.03Caption Anything in VideoCAT-V is a training-free framework for fine-grained, object-centric video captioning. It combines segmentation, temporal analysis, and multimodal prompting to produce spatially precise and temporally coherent descriptions.
Blogs
These are dated snapshots of questions I explored in 2026—not a closed research agenda. The notebook will grow across model training, behavior, interaction, embodied intelligence, robotics, and broader intelligent systems.
01Beyond Turn-Taking: Toward Models That Listen, Speak, and ActFull-duplex models, asynchronous reasoning, tools, and the two clocks of interactive intelligence.02Beyond Behavior: How Post-Training Shapes an LLM’s CharacterPost-training as behavioral specification: what SFT and reinforcement learning may shape beyond local response quality.03What Is Creativity When AI Can Generate Anything?What remains scarce when generation becomes cheap: problem choice, distant connections, and judgment.