ConvMem Turns Long-Context Reading Into a Convolution Tree and Skips the RL Training Entirely

arXiv 2609.10441·medium signalXBlueskyLinkedInCopy link

arXiv 2609.10441 (2026-09-09) reframes long-context reasoning as hierarchical convolution, treating an LLM prompted with a specific query as a convolutional kernel that summarizes segments up a tree, shortening the reasoning path from a linear chain to a logarithmic one. Unlike MemAgent-style sequential memory updating, ConvMem needs no reinforcement learning and parallelizes across both segments and reasoning threads. Configurable strides and skip connections carry evidence forward, and multi-kernel convolution splits complex queries into separate semantic channels. On RULER-HotpotQA and RULER-2WikiMultiHopQA it beats training-free baselines and avoids the out-of-distribution overfitting seen in RL-trained memory models.