ELASTIC-V: Reconstructing Spatiotemporal Fidelity via Saliency Compression and Elastic Encoding

Abstract

While Multimodal Large Language Models (MLLMs) excel in short video understanding, extending them to long videos remains challenging due to memory bottlenecks and temporal hallucinations. Existing uniform sampling and discrete pruning methods often disrupt temporal boundaries and distort physical time scales. To address this, we propose ELASTIC-V, an efficient, high-fidelity framework for long video temporal grounding. Specifically, ELASTIC-V incorporates Saliency-aware Parallel Weight Aggregation (SPWA) to adaptively suppress redundant backgrounds while preserving salient anchors under a fixed budget. Furthermore, Elastic Temporal Attribute-aware Positional Encoding (ETAPE) reconstructs temporal triplets via visual-gated Fourier mapping to restore causal consistency. We also establish CE-Bench, a multi-task benchmark evaluating hierarchical and causal complex events. Experiments demonstrate that ELASTIC-V outperforms state-of-the-art MLLMs, achieving 48.5% mIoU and an 11.5s MAE-T on CE-Bench, alongside a 5.3-point accuracy boost on Video-MME. By reducing peak VRAM by 27.3% with competitive inference latency, ELASTIC-V offers a robust, hardware-efficient paradigm for long-form video reasoning.