Debugging today's chain-of-thought reasoning for video understanding and introducing a novel reward for MLLM post-training to ground video reasoning in visual evidence.

We introduce GARE, a gap-aware retrieval framework that learns pair-specific increments to alleviate optimization tension and false-negative noise in cross-modal alignment, achieving better uniformity and semantic structure.

We propose a unified and generalizable framework for object state recognition in video data.

We propose a method to transfer video-based skill assessment knowledge across various sports domains.

We introduce a new triple-modality MLLM TriSense, which achieves comprehensive understanding of video moments by adaptively integrating visual, audio, and speech information. To support this, we propose a newly constructed dataset TriSense-2M.

CameraBench is a large-scale effort that pushes video-language models to reason about the language of camera motion just like professional cinematographers.

We introduce the task of Concept-based Video Similarity estimation, curate a dedicated benchmark and evaluate models' performance.

To alleviate the existing problems of precise timestamp dependence and limited query generalization, a weakly supervised compositional moment retrieval task is proposed, and a effective baseline is constructed to solve the new task challenges.

Introduced Refined Regularized Preference Optimization with a self-alignment framework to enable fine-grained alignment of large video language models by learning from their own errors.

DISCOVR is a self-supervised framework for echocardiography video representation learning that integrates spatial and temporal modeling, achieving strong generalization in anomaly detection, segmentation, and LVEF prediction.

Finding that today's LMMs poorly grasp the arrow of time in video, we propose ArrowRL to enhance their temporal perception and AoTBench for rigorous evaluation.

We present VideoHallu, a benchmark of over 3,000 synthetic videos with expert-crafted counterintuitive QA pairs, evaluating MLLMs' ability to detect perceptually obvious abnormalities often missed due to language priors.

HoliTom introduces a training-free holistic outer-inner token merging framework for video LLMs, significantly accelerating inference with negligible performance degradation.