arxiv arXiv cs.AI · 1h ago · research

SCOUT improves spatial reasoning in VLMs via structured CoT and process-supervised RL

from English

Researchers propose SCOUT, a framework that enhances Vision-Language Models' spatial reasoning by combining structured Chain-of-Thought with multi-objective process-reward reinforcement learning. This approach addresses credit assignment issues in existing RL methods and integrates depth perception for comprehensive 3D understanding.

SCOUT introduces a structured CoT framework to explicitly model 3D environmental perception.
The method utilizes a novel RL algorithm with multi-objective process rewards and tailored advantage estimation for fine-grained credit assignment.
A new dataset, SCOUT-24k, was synthesized through a customized pipeline to support the training.
SCOUT-3B improves upon baselines by 16.85% on general spatial benchmarks and 6.3% on complex tasks.
The larger SCOUT-7B model outperforms GPT-4o by 4.28% despite being trained exclusively on single images.

The authors consider SCOUT a critical step toward next-generation spatially-aware VLMs, demonstrating robust out-of-domain generalization to multi-image and video scenarios.

Importance 2/3 Beats a top-lab benchmark arXiv cs.AI Multimodal Research paper

Related articles

lab Microsoft Research Blog · 1d ago · 3 views

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 · 2d ago · 3 views

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.

lab NVIDIA Research · 2d ago

Hide to See introduces reasoning-prefix masking for VLM distillation

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.

arxiv arXiv cs.CL · 9d 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 · 11d 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."