lab NVIDIA Research · 15h ago · research

Hide to See introduces reasoning-prefix masking for VLM distillation

from English

Researchers propose a novel think-answer distillation framework that improves compact vision-language models' ability to utilize visual evidence by masking salient reasoning prefixes. This approach addresses the "visual forgetting" issue common in long think-answer traces, where students lose focus on visual cues during extended reasoning.

The method employs token-wise salient reasoning-prefix masking to selectively block high-influence textual cues and a self-paced masking budget scheduling strategy that scales with distillation difficulty. During training, these masks replace standard causal masks to force the student model to rely more heavily on visual information for next-token prediction.

Experimental results indicate that this approach outperforms recent open-source VLMs, VLM distillation, and self-distillation methods on multimodal reasoning benchmarks, with analysis confirming enhanced visual utilization throughout the thinking process.

Importance 1/3 Trust 3/3 NVIDIA Research Multimodal Research paper

Read original

lab Microsoft Research Blog · 4h ago · 1 view

Microsoft Research introduces CARE-X, a unified chest X-ray VLM with auxiliary supervision and tool-augmented measurement

Microsoft Research has introduced CARE-X, a research model designed to address gaps in current radiology vision-language models by combining generative report writing with calibrated diagnostic predictions. The system uses reinforcement learning (DAPO) to reward clinical correctness and pairs the Qwen3-VL-4B-Instruct model with deterministic measurement tools to evaluate performance on conditions requiring precise calculations.

lab NVIDIA Research · 15h ago · 1 view

GenRecal distills large vision-language models to small ones via recalibration

Researchers present GenRecal, a general-purpose distillation framework that transfers knowledge from large vision-language models (VLMs) to smaller, more efficient counterparts. The method addresses the challenge of heterogeneous VLM architectures by using a Recalibrator to align and adapt feature representations across different model types.

arxiv arXiv cs.CL · 8d ago · 4 views

Douyin releases DME multimodal embedding model with state-of-the-art performance

Douyin has released the technical report for its Multimodal Embedding (DME) model, which combines large-scale contrastive pre-training with latent reasoning to achieve strong discrimination and efficiency.

media Hugging Face Forums · 10d ago · 1 view

High school student seeks arXiv endorser for AttnRoute-MoE Vision Transformer paper

Hanyu (Olivia) Yu, a high school sophomore, is seeking an arXiv endorser in cs.LG or cs.CV to upload her independent preprint titled "AttnRoute-MoE: Attention-Prior Routing for Mixture-of-Experts Vision Transformers."

arxiv arXiv cs.AI · 15d ago · 1 view

ClinFusion introduces vision-centric MLLM for holistic medical understanding

Researchers introduce ClinFusion, a vision-centric multimodal large language model designed to address the challenges of deploying AI in clinical practice by unifying 2D and 3D medical image understanding. The system features a compositional cascaded vision encoder with a Cascade Spatial-Aware Locality Fusion operator and includes a new evaluation framework comprising MedIF-Bench and region-of-interest-grounded metrics.