arXiv：cs.CL· Jiangcheng Song, Haodong Jing, Yulin Mu, Lesong Tao, Jixin Wang, Jin Cui, Tianjing Ren, Runxuan Xie, Yongqiang Ma, Nanning Zheng·· 16 小时前
MemDuet：为长上下文推理学习双过程记忆
MemDuet: Learning Dual-Process Memory for Long-Context Reasoning
AI 导读
MemDuet 是一个双过程记忆框架，用小型语言模型做快速并行证据选择（System 1），用大语言模型按文档顺序做记忆更新验证与更新（System 2），分三阶段训练：System 1 监督微调、System 2 的 GRPO、双策略联合优化。
正文
Abstract:Reasoning over infinitely long documents is becoming a reality with recurrent memory agents. However, increasing context length exacerbates the trade-off between reasoning performance and inference efficiency. Inspired by Kahneman's account of fast and slow thinking, we propose MemDuet, a dual-process memory framework that learns to coordinate a small language model for fast, parallel evidence selection from raw context chunks (System 1) with a large language model for memory-update verification and memory updates in document order (System 2). We train MemDuet in three stages: supervised fine-tuning of System 1 for evidence selection, Group Relative Policy Optimization (GRPO) of System 2 for verification and memory rewriting, and joint optimization of both policies to coordinate selection and updating for long-context reasoning. Experiments on diverse long-context reasoning benchmarks (RULER-HQA, LongBench v2, and 2WikiMultiHopQA) demonstrate a favorable accuracy-efficiency trade-off as context length grows. Averaged across evaluation settings on the three datasets, MemDuet achieves an end-to-end inference speedup of 9.74x and an output-token reduction of 73.91% relative to RL-MemAgent.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.19928 [cs.CL] |
| (or arXiv:2610.19928v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.19928 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Jiangcheng Song [view email]
[v1]
Wed, 14 Oct 2026 10:16:13 UTC (576 KB)
来源：arXiv：cs.CL · arxiv.org