ReM-MoA: Reasoning Memory Sustains Mixture-of-Agents Scaling
Heng Ping, Arijit Bhattacharjee, Peiyu Zhang, Shixuan Li, Wei Yang, Ali Jannesari, Nesreen Ahmed, Paul Bogdan
Why It Matters
What makes this one worth your time
This work is relevant for AI researchers focused on improving the scalability and efficiency of multi-agent systems, particularly in reasoning tasks.
ReM-MoA enhances Mixture-of-Agents scaling with memory-augmented reasoning mechanisms.
Summary
The paper introduces ReM-MoA, a memory-augmented framework for Mixture-of-Agents architectures that aims to sustain scaling benefits as depth increases. It employs a Ranked Reasoning Memory and a Curated Diversified Memory Routing scheme to enhance reasoning trace storage and agent exposure, respectively. The framework is tested across five reasoning benchmarks, showing improved performance over existing MoA variants.
Key contributions
- Introduction of a memory-augmented MoA framework.
- Development of a Ranked Reasoning Memory for cross-layer reasoning.
- Implementation of a Curated Diversified Memory Routing scheme.
Notable insights
- The use of a Ranked Reasoning Memory to store and rank reasoning traces across layers.
- A Curated Diversified Memory Routing scheme that maintains exploration diversity while enhancing reasoning quality.
Possible limitations
- Not stated in the abstract
Abstract
arXiv:2606.24437v1 Announce Type: new Abstract: Mixture-of-Agents (MoA) architectures improve inference-time scaling by organizing multiple LLM agents into layered reasoning pipelines. However, existing MoA variants fail to sustain gains as depth increases, exhibiting degradation, early plateauing, or saturation. We propose ReM-MoA, a memory-augmented MoA framework that sustains scaling through two mechanisms: (1) a Ranked Reasoning Memory that persistently stores and ranks reasoning traces from all layers using a comparative Reviewer Agent, and (2) a Curated Diversified Memory Routing scheme that exposes different agents to distinct combinations of successful and failed traces, preserving exploration diversity while propagating high-quality reasoning. We further introduce an optional multi-domain Reviewer distillation pipeline that improves ranking quality through frontier-model supervision. Across five reasoning benchmarks spanning math, formal logic, code, knowledge, and commonsense, ReM-MoA consistently outperforms prior MoA variants across both depth and width scaling, and its advantage widens with depth, establishing structured cross-layer reasoning memory as a key missing mechanism for scalable multi-agent inference.