Multimodal LLMs may rely on language priors rather than pertinent visual evidence, especially on long documents. I explore agentic perception frameworks that gather evidence iteratively to improve accuracy and efficiency.

• InSight-doc: Agentic Visual Perception for Long-Document Understanding
  In submission, 2026
  Replaces fixed-resolution, single-pass pipelines with iterative perception that selectively acquires high-resolution crops on demand.
  Paper, code, and data will be publicly available soon.

Controllable editing requires precise spatial and semantic guidance without costly retraining. I develop training-free methods that combine structural control with flexible prompt guidance.

• CannyEdit: Selective Canny Control and Dual-Prompt Guidance for Training-Free Image Editing
  IEEE CAI, 2026
  Enables selective edge-based structural control with dual-prompt guidance for training-free, controllable image editing.
  [Paper] [Code]

Foundation models can lose robustness during fine-tuning and fail under distribution shift. I design training objectives that anchor decisions to invariant, generalizable features.

• Consistency Regularization for Domain Generalization with Logit Attribution Matching
  UAI, 2024
  Logit Attribution Matching (LAM) anchors predictions to domain-invariant causal features.
  [Paper] [Code]
• Dual Risk Minimization: Towards Next-Level Robustness in Fine-Tuning Zero-Shot Models
  NeurIPS, 2024
  Combats robustness vanishing during foundation-model fine-tuning via dual risk minimization.
  [Paper] [Code]