Shu-Lin Tian

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.

Evaluation Agent slashes evaluation time to 10% of traditional methods while providing detailed, user-tailored analyses of visual generative models.

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

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

Turns out, just showing a vision-language model the last few frames of a video stream beats fancy architectures designed for long-term memory.

Today's best AI agents can only achieve 48% accuracy when reasoning about your personal files, revealing a surprising gap between lab performance and real-world usability.