![[2602.14236] Dual-Signal Adaptive KV-Cache Optimization for Long-Form Video Understanding in Vision-Language Models](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.14236] Dual-Signal Adaptive KV-Cache Optimization for Long-Form Video Understanding in Vision-Language Models
Summary
The paper presents Sali-Cache, a novel optimization framework for Vision-Language Models (VLMs) that addresses memory bottlenecks in long-form video processing by implementing dual-signal adaptive caching.
Why It Matters
As video content continues to grow in complexity and length, optimizing memory usage in Vision-Language Models is crucial for enhancing performance and accessibility. Sali-Cache's proactive memory management can significantly improve efficiency, making advanced video processing feasible on consumer-grade hardware.
Key Takeaways
- Sali-Cache optimizes memory usage in VLMs by implementing dual-signal adaptive caching.
- The framework uses optical flow and saliency detection to manage memory allocation proactively.
- Achieves a 2.20x compression ratio in effective memory usage while maintaining accuracy.
- Preserves context-rich features over longer durations without degrading performance.
- Enables efficient processing of long-form video content on consumer-grade hardware.
Computer Science > Computer Vision and Pattern Recognition arXiv:2602.14236 (cs) [Submitted on 15 Feb 2026] Title:Dual-Signal Adaptive KV-Cache Optimization for Long-Form Video Understanding in Vision-Language Models Authors:Vishnu Sai, Dheeraj Sai, Srinath B, Girish Varma, Priyesh Shukla View a PDF of the paper titled Dual-Signal Adaptive KV-Cache Optimization for Long-Form Video Understanding in Vision-Language Models, by Vishnu Sai and 4 other authors View PDF HTML (experimental) Abstract:Vision-Language Models (VLMs) face a critical memory bottleneck when processing long-form video content due to the linear growth of the Key-Value (KV) cache with sequence length. Existing solutions predominantly employ reactive eviction strategies that compute full attention matrices before discarding tokens, resulting in substantial computational waste. We propose Sali-Cache, a novel a priori optimization framework that implements dual-signal adaptive caching through proactive memory management. By integrating a temporal filter based on optical flow analysis for detecting inter-frame redundancy and a spatial filter leveraging saliency detection for identifying visually significant regions, Sali-Cache intelligently manages memory allocation before entering computationally expensive attention operations. Experimental evaluation on the LLaVA 1.6 architecture demonstrates that our method achieves a 2.20x compression ratio in effective memory usage while maintaining 100% accuracy across BL...