Research Focus
Visual Memory Systems for AI Agents
I study AI agents that can organize useful memories, understand visual evidence, and improve through post-training. These three directions share one goal: building agents that remember what matters and reason reliably before they act.
01 · Memory Systems
How can agents organize and reuse experience?
CLAG (ACL 2026 Findings) augments AI agents with self-organizing memory, structuring past experiences for more focused retrieval and reuse.
02 · Visual Understanding
How can agents reason from what they see?
BUZZY (Preprint, 2026) uses contrastive scoring to suppress text-induced choice bias and ground answers in visual evidence without additional training. Breaking Failure Cascades (EMNLP 2026 Main) extends this direction to medical multimodal reasoning.
03 · Post-Training for Agents
How can agents learn to plan and use tools?
Distilling Expert-level Planning (ICML Workshop 2026) uses reinforcement learning to improve tool use for real-world diabetes prescribing. In Breaking Failure Cascades (EMNLP 2026 Main), step-aware reinforcement learning is tailored to multimodal reasoning failures.