C2-Routing cuts VLM visual tokens by routing crop or compress
With a 1,024-token cap already applied, the method reports 36.53% fewer FLOPs and a 0.77% mean relative score drop.
Stories per day · 30 days
4 RSS
Relatedmultimodal6robotics6spatial-reasoning4agents3ocr3open-weights3qwen3reinforcement-learning3benchmark2benchmarks2gui-grounding2nemotron2nvidia2object-detection2retrieval2vlm2world-models23d-gaussian-splatting1
74 stories, newest first · one card per story
Stories tagged #vision-language
Study rechecks where Platonic representation alignment appears in newer models
The capability-alignment trend holds when captions are clean, but same-modality alignment still far exceeds cross-modal scores.
SLA Model Scales Spatial Understanding via Continual Pre-training
SLA optimizes action prediction and spatial alignment from frozen 3D perception for robotic manipulation.
VFlash Enables Fast Visual Planning via Flash Primitive Decoding
VFlash predicts visual primitives jointly for 25x faster generation in robot trajectory tasks.
HALO synthesizes vision-language-motion data for humanoid robots
HALO achieves 91 percent success on composite tasks in unmapped real homes by scaling simulation data rather than human teleoperation.
DGP boosts MLLM performance via dynamic gated pathways
Dynamic Gated Pathways connect all ViT and LLM layers to improve multimodal understanding in LLaVA-1.5 models.
DR Pixel enables page-native deep research with vision-language models
DR Pixel is an open-source agent that reasons directly over document page images, outperforming OCR-based methods on scientific papers.
Safety4DM vision-language framework for risk-aware dynamic maps in collaborative driving
Fine-tuned Qwen3-VL-4B reaches 89.83% accuracy on RoadRisk dataset with SafeStep interface achieving over 93% query accuracy.
VLSplat uses vision-language for object-centric 3D Gaussian Splatting
VLSplat augments 2D-to-3D semantic lifting with vision-language model priors and scene graph refinement to improve object-level semantic consistency and structural completeness.
Dwell lets vision-language agents learn policies in latent space
The framework treats a pretrained VLM as a recurrent latent transition model and reports top trainable results on five visual environments with less imagination compute.
GUI-JEPA scores mobile GUI actions in latent space without decoding
On AndroidControl it is 2.8–4.0x faster than a fine-tuned Qwen2.5-VL-7B agent while using 6% of the parameters.
AdaGround adds adaptive zoom and a verifier for GUI grounding
On Qwen3-VL, the training-free method raises accuracy 7.4 to 20.6 points on OSWorld-G and ScreenSpot-Pro.
EvidenceConnect data training produces ZwZ2 multimodal perception gains
Zoomed-region questions plus RLVR improved spatial, long-context, agent, and general visual reasoning without benchmark-specific data.
Cross-modal steering vectors transfer between text and image generation
In unified multimodal models, vectors found on one modality steer the other zero-shot, including concepts unique to one side.
RoboJev-27B open model scores robot manipulation success from video
On 268 held-out SO-101 episodes the authors report 95.2% balanced accuracy and 0.993 AUROC.
PreView improves GUI grounding without retraining the model
A lightweight preprocessor adds zoomed views and cues, lifting accuracy for three frozen vision-language models on five benchmarks.
HAD-VLM pairs diffusion with language for ultrasound measurements
It cuts four EchoNet-LVH measurement errors 21.4–37.6% versus a Vicuna baseline and lowers X-ray vertebral error from 11.4 to 6.7 pixels.
Paper introduces Ψ-0.6 vision-language world model with streams of thought
The authors say Ψ-0.6 beats leading open-weight models zero-shot on multi-view spatial reasoning and can simulate counterfactual scenes.
WeVisOne-VL unifies object- and pixel-level visual perception
A pixel decoder behind an <OBJ> token emits boxes, masks, and mattes after curriculum SFT and mixed-task RL.
SpatialBrain stores spatial-reasoning experience with a fixed backbone
On five benchmarks it beats direct answers from the same backbone on four and ties on the fifth.
OvisOCR3, a 0.8B model, covers five OCR task families
Weights are on Hugging Face; authors report 96.74 on OmniDocBench v1.6 and first place on OCRBench v2.
General VLMs rerank multimodal hits with one choice question
Zero-shot Qwen3.8-27B beats a supervised 2B reranker on three benchmarks and matches Reranker-8B on visual documents.
Liquid AI releases LFM3-VL-2B on-device vision-language model
The 1.92B model takes text, images, and video, reasons step by step, and uses the LFM Open License v1.0.
1
MMedAgent-Plus coordinates 18 specialist models across medical tasks
End-to-end final-answer accuracy is 0.98, matching the dedicated specialist and above a 0.28 baseline.