Yuhao Dong

Most multimodal models fail to truly understand tool use in real-world egocentric videos, with even Gemini-3.1-Pro struggling to ground actions in visual evidence.

This work builds Video-HopChain, a dataset of 22,550 multi-hop video questions over 13,378 videos, together with a held-out benchmark of 1,000 questions, which raises the mean over eight video understanding and reasoning benchmarks from 55.4 to 57.9 and introduces Confidence-Gated Exploration (CGE).

Rubric-based scoring reveals that vision-language models can significantly improve their grounding in visual evidence, enhancing reasoning and instruction adherence.

Atomic visual perception in MLLMs is largely unsolved, with no model surpassing 60% accuracy on a new benchmark designed to isolate perceptual capabilities.

Kimi K3's innovative architecture achieves a 2.5x scaling efficiency improvement, enabling robust performance across diverse long-horizon tasks.

Current video generation models struggle with law-grounded reasoning, with the best achieving only 47% on the new Apple-PI benchmark.

VideoChat3 achieves unprecedented generalization in video understanding while maintaining high efficiency, outperforming larger models with just 4 billion parameters.

ViQ achieves a groundbreaking balance between semantic richness and detail in visual representations, enabling efficient multimodal training without sacrificing quality.

Spatial reasoning can be transformed from isolated frame predictions to dynamic scene understanding, significantly boosting performance in multi-view and video tasks.

Ditching modular architectures unlocks surprisingly competitive vision-language performance, proving that end-to-end pixel-to-word models can rival traditional approaches at scale.

Current audio-visual generation models struggle to maintain coherence and alignment when scaling to minute-long content, a problem exposed by the new LongAV-Compass benchmark.

Leaderboard-topping video models are still surprisingly brittle, failing on basic video reasoning tasks unless given the right textual cues.

Forget dialogue summaries – FileGram builds user profiles directly from atomic file-system actions, unlocking a richer, more privacy-preserving approach to agent personalization.

Today's best multimodal models can only solve half of compositional visual tool-use tasks, revealing a critical gap in their ability to plan and execute complex, multi-step visual reasoning.