Causality Matters: How Temporal Information Emerges in Video Language Models

Yumeng Shi; Quanyu Long; Yin Wu; Wenya Wang

Home/Archive/arXiv

arXiv · 2508.11576

Causality Matters: How Temporal Information Emerges in Video Language Models

Yumeng Shi·Quanyu Long·Yin Wu·Wenya Wang

View original source ↗Read source document ↗Export metadata ↓

Abstract

Video language models (VideoLMs) have made significant progress in multimodal understanding. However, temporal understanding, which involves identifying event order, duration, and relationships across time, still remains a core challenge. Prior works emphasize positional encodings (PEs) as a key mechanism for encoding temporal structure. Surprisingly, we find that removing or modifying PEs in video inputs yields minimal degradation in the performance of temporal understanding. In contrast, reversing the frame sequence while preserving the original PEs causes a substantial drop. To explain this behavior, we conduct substantial analysis experiments to trace how temporal information is integrated within the model. We uncover a causal information pathway: temporal cues are progressively synthesized through inter-frame attention, aggregated in the final frame, and subsequently integrated into the query tokens. This emergent mechanism shows that temporal reasoning emerges from inter-visual token interactions under the constraints of causal attention, which implicitly encodes temporal structure. Based on these insights, we propose two efficiency-oriented strategies: staged cross-modal attention and a temporal exit mechanism for early token truncation. Experiments on two benchmarks validate the effectiveness of both approaches. To the best of our knowledge, this is the first systematic study of video temporal understanding in VideoLMs, offering insights for future model improvement. Our code is available at https://github.com/ANDgate99/Causality-Matters .

Explore related subjects

cs.CV

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Yumeng Shi, Quanyu Long, Yin Wu, Wenya Wang. 2025-11-15. Causality Matters: How Temporal Information Emerges in Video Language Models. https://arxiv.org/abs/2508.11576

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Band-Attention Modulation Network for Robust Face Forgery Detection

Face forgery detection faces critical challenges in generalizing to unseen manipulation techniques and remaining robust under image compression, which often obscures subtle artifacts. Existing methods typically rely on fixed filters or coarse band separation, lacking the adaptability to learn task-specific spectral cues. To address this, we propose the Band-Attention Modulation Network (BAM-Net), a novel framework that pioneers learnable, fine-grained modulation of frequency components for forgery detection. At its core is the Band-Attention Modulation (BAM) mechanism, which transforms an image into its Discrete Cosine Transform (DCT) spectrogram and learns to dynamically reweight frequency bands along anti-diagonals. This process effectively enhances forgery-related spectral signatures while suppressing less informative ones, simulating an adaptive "inverse compression" that counters information loss. The modulated frequency information is then fused with the spatial domain to guide a lightweight yet effective spatial backbone equipped with distance-decayed attention for comprehensive feature extraction. Extensive experiments on FaceForensics++, Celeb-DF, and DFDC datasets demonstrate that BAM-Net achieves state-of-the-art performance. More importantly, it exhibits exceptional generalization in cross-dataset, cross-compression, and cross-manipulation scenarios, underscoring the vital role of adaptive frequency band modulation in building robust forgery detectors.

cs.CV↗

arXiv2026

Comparing YOLOv11 and YOLOv8 for instance segmentation of occluded and non-occluded immature green fruits in complex orchard environment

This study conducted a comprehensive performance evaluation on YOLO11 (or YOLOv11) and YOLOv8, the latest in the "You Only Look Once" (YOLO) series, focusing on their instance segmentation capabilities for immature green apples in orchard environments. YOLO11n-seg achieved the highest mask precision across all categories with a notable score of 0.831, highlighting its effectiveness in fruit detection. YOLO11m-seg and YOLO11l-seg excelled in non-occluded and occluded fruitlet segmentation with scores of 0.851 and 0.829, respectively. Additionally, YOLOv11x-seg led in mask recall for all categories, achieving a score of 0.815, with YOLO11m-seg performing best for non-occluded immature green fruitlets at 0.858 and YOLOv8x-seg leading the occluded category with 0.800. In terms of mean average precision at a 50\% intersection over union (mAP@50), YOLOv11m-seg consistently outperformed, registering the highest scores for both box and mask segmentation, at 0.876 and 0.860 for the "All" class and 0.908 and 0.909 for non-occluded immature fruitlets, respectively. YOLO11l-seg and YOLOv8l-seg shared the top box mAP@50 for occluded immature fruitlets at 0.847, while YOLO11m-seg achieved the highest mask mAP@50 of 0.810. Despite the advancements in YOLO11, YOLOv8n surpassed its counterparts in image processing speed, with an impressive inference speed of 3.3 milliseconds, compared to the fastest YOLO11 series model at 4.8 milliseconds, underscoring its suitability for real-time agricultural applications related to complex green fruit environments. Future work will compare YOLO26 (YOLOv26) and YOLO27 (YOLOv27) using the same dataset and training protocol.

cs.CV↗

arXiv2026

Cross-Task Generalization in Handwriting-Based Alzheimer's Screening via Vision Language Adaptation

Alzheimer's disease (AD) is a prevalent neurodegenerative disorder for which early detection is critical. Handwriting, which can be disrupted by subtle motor and cognitive decline, provides a non-invasive and cost-effective window for AD screening. Existing handwriting-based AD studies mostly rely on online trajectories and hand-crafted features, while the influence of handwriting task type on diagnostic performance and cross-task generalization remains underexplored. Meanwhile, large-scale vision--language models have demonstrated strong transfer and adaptation ability in natural-image anomaly detection and several medical modalities, such as chest X-ray and brain MRI. However, handwriting-based disease detection remains unexplored within this paradigm. To address this gap, we introduce a lightweight Cross-Layer Fusion Adapter (CLFA) framework that repurposes Contrastive Language--Image Pre-training (CLIP) for handwriting-based AD screening. CLFA inserts multi-level adapters into a frozen visual encoder, combining cross-layer feature fusion with depthwise 2D convolution on patch grids to capture both local stroke irregularities and higher-level handwriting structure. This design progressively aligns pretrained vision--language representations with AD-related handwriting cues and supports transfer from supervised source tasks to task-disjoint unseen target tasks. On the Darwin dataset, under the subject-disjoint cross-task protocol, averaged over all 600 task-disjoint source-target pairs, CLFA achieves 74.63\% AUC, 74.85\% accuracy, and 73.72\% F1 score, outperforming the best competing model by 2.15, 1.79, and 1.87 percentage points, respectively.

cs.CV↗