Tag
This paper introduces Evidence-Backed Video Question Answering (E-VQA), a new task requiring models to output both semantic answers and precise spatio-temporal evidence like tracked object segmentation masklets. The authors create a human-verified benchmark and a scalable training dataset, showing significant improvements over baselines.
GUICrafter introduces a weakly-supervised GUI agent that leverages massive unannotated screenshots and a two-stage curriculum learning framework to reduce reliance on expensive human annotations, achieving competitive performance with advanced systems like UI-TARS using only 0.1% of its data.
NVIDIA has open-sourced the visual grounding model LocateAnything-3B, which can accurately detect and bound all target objects in dense scenes.
NVIDIA released LocateAnything, an open-source model that achieves ~10x faster object detection by predicting all coordinates simultaneously instead of sequentially, reaching 12.7 FPS on a single H100 and outperforming 32B parameter models.
Introduces iVGR, a reinforcement learning framework that internalizes visual localization into textual reasoning for multimodal language models, eliminating the need for explicit visual grounding during inference while improving fine-grained perception performance.
NVIDIA introduces LocateAnything, a unified generative grounding and detection framework that uses Parallel Box Decoding to improve decoding throughput and localization accuracy. This work will be presented at CVPR 2026.
LocateAnything proposes Parallel Box Decoding for unified visual grounding and object detection, decoding geometric elements as atomic units to improve throughput and localization accuracy, supported by a large-scale dataset of 138M samples.
This paper presents a function-centric framework using Transcoders to trace computational pathways in vision-language models, demonstrating stronger attribution of visual grounding and the ability to predict hallucinations via graph-based features.
This paper introduces ForMaT, a parallel corpus of 3,956 PDFs across 15 language pairs designed for visually-grounded multilingual translation, preserving layout metadata to benchmark layout-aware MT systems.
MementoGUI introduces a plug-in agentic memory framework for GUI agents that uses learned controllers for selective memory management and retrieval, improving performance on long-horizon tasks with compressed visual and textual representations.
This paper presents a case study on visually-guided movie subtitle translation for low-resource Indic languages, demonstrating that selective visual grounding improves translation quality while addressing temporal misalignment challenges.
The paper introduces SeePhys Pro, a benchmark to diagnose modality transfer issues in multimodal RL for physics reasoning, revealing that models struggle with representation-invariant reasoning and often rely on residual textual cues rather than visual evidence.
HyperEyes is a parallel multimodal search agent that uses dual-grained reinforcement learning to optimize inference efficiency, achieving higher accuracy with significantly fewer tool-call rounds compared to existing agents.