ConvMem: Convolutional Memory for Long-Context Reasoning
ConvMem turns long-context reasoning into a parallel hierarchy instead of a step-by-step memory chain.
The paper proposes a training-free framework that has an LLM summarize text segments like a convolutional kernel. Its tree structure is meant to shorten the reasoning path and reduce latency versus sequential memory methods such as MemAgent. The authors add configurable strides, skip connections, and multi-kernel convolution to preserve evidence and split complex queries into semantic channels. Experiments on RULER-HotpotQA and RULER-2WikiMultiHopQA show gains over training-free baselines and fewer out-of-distribution issues than RL-trained memory models. ArXiv · AI/CL/LG's note
The paper proposes a training-free framework that has an LLM summarize text segments like a convolutional kernel. Its tree structure is meant to shorten the reasoning path and reduce latency versus sequential memory methods such as MemAgent. The authors add configurable strides, skip connections, and multi-kernel convolution to preserve evidence and split complex queries into semantic channels. Experiments on RULER-HotpotQA and RULER-2WikiMultiHopQA show gains over training-free baselines and fewer out-of-distribution issues than RL-trained memory models. ArXiv · AI/CL/LG's note
score 5