Context compression finally works in production: new research cuts LLM input 16x without the accuracy hit

LCLMs compress LLM context before decode — 8.8x faster at 16x compression, beating every KV cache method tested. Open-sourced by NYU and Columbia.

Open-source FLIM Playground could speed reproducible analysis of complex cell images

Modern fluorescence microscopy can generate images of living cells as stunning to look at as they are informative to study. For techniques like fluorescence lifetime imaging microscopy (FLIM), those ...
IEEE

Discretized Gaussian Mixture Hyperprior for Learned Image Compression with Mask Module

Abstract: Learned image compression approaches have shown great potential with promising results. However, according to the commonly used measurement methods, there still lies a performance gap ...