MemDuet: Learning Dual-Process Memory for Long-Context Reasoning

Key points

• Reasoning over infinitely long documents is becoming a reality with recurrent memory agents.1
• However, increasing context length exacerbates the trade-off between reasoning performance and inference efficiency.1
• 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.1
• Experiments on diverse long-context reasoning benchmarks (RULER-HQA, LongBench v2, and 2WikiMultiHopQA) demonstrate a favorable accuracy-efficiency trade-off as context length grows.1

Sources (1)

• [1]MemDuet: Learning Dual-Process Memory for Long-Context Reasoning
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 10:16 AM

  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).

  Reasoning over infinitely long documents is becoming a reality with recurrent memory agents.

Extractive summary: sentences quoted from the sources.