Tiny vision-language models for visual document retrieval, built on UltraBERT.
AI & ML interests
Late interaction and multi-vector retrieval: ColBERT-style text retrievers, ColPali-style visual document retrieval, MaxSim scoring, token pruning and index compression, efficient training recipes, and open benchmarks.
Recent Activity
One LoRA switches a vision-language model between late-interaction retrieval and generation.
Training and evaluation data for visual document retrievers.
Demos of the UltraVBERT, ColQwen3.5 and Hydra models.
Late-interaction visual document retrievers on Qwen3.5.
Korean visual document retrievers on Qwen3-VL-Embedding-2B.
Papers on late-interaction and visual document retrieval.
-
Hydra: Unifying Document Retrieval and Generation in a Single Vision-Language Model
Paper • 2603.28554 • Published -
Spatially-Grounded Document Retrieval via Patch-to-Region Relevance Propagation
Paper • 2512.02660 • Published • 1 -
KoVRE: Training an Efficient Embedding Model for Korean Visual Document Retrieval
Paper • 2608.01389 • Published
Tiny vision-language models for visual document retrieval, built on UltraBERT.
Late-interaction visual document retrievers on Qwen3.5.
One LoRA switches a vision-language model between late-interaction retrieval and generation.
Korean visual document retrievers on Qwen3-VL-Embedding-2B.
Training and evaluation data for visual document retrievers.
Papers on late-interaction and visual document retrieval.
-
Hydra: Unifying Document Retrieval and Generation in a Single Vision-Language Model
Paper • 2603.28554 • Published -
Spatially-Grounded Document Retrieval via Patch-to-Region Relevance Propagation
Paper • 2512.02660 • Published • 1 -
KoVRE: Training an Efficient Embedding Model for Korean Visual Document Retrieval
Paper • 2608.01389 • Published
Demos of the UltraVBERT, ColQwen3.5 and Hydra models.