TopoCompress: Long Context Compression via Graph-Wired Semantic Trajectories

Daniel Agyei Asante, Yang Li

Long-context compression is essential for reducing the cost and latency of large language model inference. However, existing methods can fragment important evidence, require additi…