MEGA Hub

PI-Mem: Pushing Long-Context Reasoning to 3.6M Tokens with Parallel-Iterative Memory

Authors

Do you know Dawei Liu?You can claim authorship or link another user.Do you know Haixu Song?You can claim authorship or link another user.Do you know Shuang Cheng?You can claim authorship or link another user.Do you know Shijie Wang?You can claim authorship or link another user.Do you know Haozheng Hou?You can claim authorship or link another user.Do you know Kaifeng Liu?You can claim authorship or link another user.Do you know Ermo Hua?You can claim authorship or link another user.Do you know Zhonghang Yuan?You can claim authorship or link another user.Do you know Zhijie Zhong?You can claim authorship or link another user.Do you know Yuchen Fan?You can claim authorship or link another user.Do you know Biqing Qi?You can claim authorship or link another user.Do you know Bowen Zhou?You can claim authorship or link another user.

Abstract

Long-context reasoning remains a critical bottleneck for large language models, as recent recurrent-memory approaches face two inherent challenges: sequential chunk-wise updates can overwrite early critical evidence with later irrelevant content, and serial inter-chunk dependencies limit parallelism and cause latency to increase with context length. To address these issues, we propose PI-Mem (Parallel-Iterative Memory), a mechanism that processes all chunks in parallel and iteratively refines a shared memory over a bounded number of turns. In each turn, PI-Mem reads all chunks in parallel conditioned on the current memory, selects new or complementary evidence from each chunk, and merges the selected evidence into a compact shared memory for the next turn. To discourage redundant turns, we optimize the workflow through reinforcement learning with an auxiliary turn-efficiency reward, enabling the model to adaptively exit once sufficient evidence has been accumulated. We evaluate PI-Mem with Qwen3.5-35B-A3B and Qwen2.5-7B on the HotpotQA benchmark across context lengths up to 3.6 million tokens and find that it outperforms the recurrent-memory baseline by +6.25 and +7.81 absolute points while achieving 6.1$\times$ and 2.1$\times$ inference speedups, respectively. These results demonstrate that PI-Mem breaks the accuracy--efficiency trade-off in long-context reasoning and provides a scalable approach to complex multi-hop question answering over extremely long documents.

Community

00