Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 26·5 min read
ST-Lite: Training-Free KV Cache Compression with Spatio-Trajectory Guidance for Long-Horizon GUI Agents
TL;DR
ST-Lite is a training-free compression scheme designed for key-value (KV) caches in GUI agents, achieving up to 2.35x decoding speedup while maintaining accuracy. It effectively addresses unique GUI interaction properties, outperforming existing methods in memory-constrained environments.
✦ Why It Matters
Implement ST-Lite to enhance the performance of your GUI agents under memory constraints today.
Key Takeaways
- Utilize ST-Lite for efficient KV cache compression in GUI applications.
- Leverage Trajectory-aware Semantic Gating to reduce redundant data.
- Apply Component-centric Spatial Saliency for better UI element preservation.
- Optimize resource allocation with a flat per-layer budget approach.
How It Works
ST-Lite employs three key strategies to enhance KV cache compression for GUI agents. First, Trajectory-aware Semantic Gating (TSG) identifies and filters out redundant historical frames, reducing unnecessary data.
Second, Component-centric Spatial Saliency (CSS) focuses on preserving the boundaries of UI elements, ensuring that important visual information is retained. Finally, the flat per-layer budget approach prevents misallocation of resources across different layers, optimizing the overall performance of the compression scheme.