RegionReasoner: Region-Grounded Multi-Round Visual Reasoning
RegionReasoner improves multi-round visual reasoning by grounding each reasoning trace in explicit regions and aligning local visual details with global scene semantics.
Updates
News
- Selected as a finalist for the 2026 Qualcomm Innovation Fellowship for Europe.
- One paper was accepted by ICLR 2026.
- One paper was accepted by WACV 2026.
- One paper was accepted by NeurIPS 2025.
- One paper was accepted by NeurIPS 2024.
- One paper was accepted by CVPR 2024 Workshop.
- One paper was accepted by ICML 2023.
Selected Publications
Research
QUOTA: Quantifying Objects with Text-to-Image Models for Any Domain
QUOTA uses a domain-agnostic optimization framework to improve object-count control for text-to-image models across unseen domains without retraining.
The Curse of Depth in Large Language Models
LayerNorm Scaling mitigates variance explosion in deep Transformer layers, helping large language models benefit more consistently from depth.
IPO: Interpretable Prompt Optimization for Vision-Language Models
IPO uses language and multimodal models to generate dataset-specific, human-readable prompts that improve generalization for vision-language models.
Training-Free Semantic Segmentation via LLM-Supervision
A training-free semantic segmentation framework that uses large language models to build richer class descriptors and ensemble subclass-level predictions.
MetaModulation: Learning Variational Feature Hierarchies for Few-Shot Learning with Fewer Tasks
MetaModulation increases task diversity in few-shot learning by modulating feature hierarchies, including variational variants for task uncertainty.