Preprint describes KV-cache compression for GUI agents

Researchers introduced ST-Lite, a training-free KV-cache compression method for vision-language GUI agents, in an arXiv preprint. It filters redundant interface frames, retains UI-element boundaries and uses a flat per-layer cache budget. Across seven GUI benchmarks and two backbones, the authors report matching or exceeding full-cache task accuracy on the primary backbone at a 20% budget, with up to 2.35x decoding speedup at fivefold compression. The implementation is linked from the preprint.