Gated-Memory Routing Boosts Multi-Agent LLM Efficiency and Accuracy

Multi-agent Large Language Model (LLM) systems are designed to tackle complex reasoning tasks by coordinating several agents. A key challenge in these systems is adapting the orchestration to the evolving state of collaboration. Relying solely on the initial query for routing decisions can lead to inaccuracies because it ignores intermediate progress or errors. Conversely, using the complete execution history provides context but creates an "execution-history overload," as later decisions must process every prior step, including redundant ones, significantly increasing computational cost. To address this, researchers propose Gated-Memory Routing. This framework conditions each decision on the initial query and a dynamically learned execution memory. It incorporates a "Memory Write Gate" that selectively commits only non-redundant reasoning steps to memory, ensuring a compact and informative state. Additionally, a "Retrieval Gate" provides each agent with only the most relevant subset of this memory, preventing the accumulation of unnecessary context. At each step, the system intelligently selects the next agent role and backbone based on this refined memory. An Adaptive Halting Controller then stops execution once sufficient evidence for an answer is present in the memory. Across five reasoning and code-generation benchmarks, Gated-Memory Routing demonstrated superior performance, achieving the best average accuracy (exceeding baselines by 2.44 points) while significantly reducing inference costs (e.g., 31.9% reduction on HumanEval).