ReM-MoA: Reasoning Memory Sustains Mixture-of-Agents Scaling

2026-06-23

AI Summary

This work addresses the performance degradation and premature saturation observed in existing Mixture-of-Agents (MoA) architectures as reasoning depth increases. To overcome these limitations, the authors propose a memory-augmented hierarchical reasoning framework that incorporates Ranked Reasoning Memory to establish structured, persistent cross-layer memory. A Curated Diversified Memory Routing mechanism is introduced to enable efficient memory scheduling. Additionally, a comparative Reviewer Agent ranks reasoning trajectories and facilitates multi-domain knowledge distillation. Evaluated across five benchmarks—mathematics, formal logic, code generation, factual knowledge, and commonsense reasoning—the proposed method consistently outperforms current MoA approaches as both depth and width scale, with performance gains becoming markedly more pronounced at greater reasoning depths.

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.

Problem

• Mixture-of-Agents
• scaling
• reasoning
• memory
• LLM

Innovation

• Reasoning Memory
• Mixture-of-Agents
• Memory Routing
• Reviewer Agent
• Inference Scaling