Long-form video understanding is crucial for tasks such as movie comprehension, where video question answering (videoQA) serves as a core capability. Existing videoQA systems struggle with temporally scattered evidence and cross-modal misalignment; models fine-tuned on short clips often degrade sharply on long-form benchmarks due to fragmented evidence reasoning. We propose TCRS-QA, a multi-stage, training-free framework that unifies fine-grained perception with iterative reasoning by treating each shot as the atomic unit. TCRS-QA integrates four complementary modules: (i) hierarchical evidence retrieval for diverse keyframe selection, (ii) semantic alignment and textual grounding to filter shot-level descriptions, (iii) narrative rhythm detection to identify narrative phase transitions, and (iv) iterative chain-of-thought reasoning for multimodal inference and verification. On the ICCV 2025 SF20K Story-Level Movie Understanding Audio Description Challenge, TCRS-QA achieves state-of-the-art performance, and delivers superior performance on Video-MME and LongVideoBench datasets.

Video Anomaly Understanding (VAU) extends traditional Video Anomaly Detection by enabling both identification and interpretation of abnormal events via natural language. Building on instruction-tuned multimodal large language models (MLLMs), we present a unified VAU framework and perform multi-task LoRA fine-tuning. We construct PreVAD-Instruct27k, a large-scale dataset of 27,854 videos annotated with categories, descriptions, and open-ended question answering (QA) pairs. PreVAD-Instruct27k enables joint optimization of classification, description, and QA tasks. Under sparse frame sampling, our model achieves 77.72% classification accuracy, 93.62% UniEval score for abnormal description, and 90.64% UniEval score for QA on PreVAD-Instruct27k, while maintaining strong generalization on XD-Violence and UCF-Crime. Ablation studies confirm the effectiveness of the multi-task design.

The aim of the violent recognition task is to determine whether a video contains violent behaviors. Given that violent behavior often comes with visual and audio anomalies, multimodal approaches have always played an important role in this field. However, existing methods have been limited by the insufficient utilization of audio-visual self-supervised semantic cues and correlation, resulting in a restricted representational capacity of the network and low generalization due to the scarcity of available violent video datasets. To address this issue, we propose a violent action recognition model based on global-local visual and audio contrastive learning. Our model introduces global and local contrastive objectives to achieve audio-visual multi-grained semantic alignment and leverage the correlation for violent video recognition. Experimental results demonstrate that our proposed model improves state-of-the-art by 2.31% on the VSD dataset, 0.71% on the Violent-Flows dataset, and 1.43% on the VCD dataset.