ATOM framework boosts multi-agent AI efficiency by 30% with budget control

Large language model-based multi-agent systems often struggle with the stability-extensibility trade-off and mismatched computational budgets. To address this, a team of researchers proposes ATOM, an adaptive framework that generates budget-controllable collaboration graphs via task-driven reinforcement learning. Inspired by atomic structures, ATOM employs a nucleus-electron hierarchy: a stable, offline-learned collaboration backbone (nucleus) persists across tasks, while query-conditioned agents (electrons) are dynamically activated during inference based on query difficulty.

A key innovation is the complexity-aware budgeting strategy, which estimates query difficulty to strictly regulate electron instantiation, ensuring resource consumption aligns with task demands. This prevents over-allocating compute to simple queries or under-provisioning complex ones. Across six diverse benchmarks—including reasoning, code generation, and question answering—ATOM achieves state-of-the-art performance while improving token efficiency by up to 30% compared to strong baselines. The framework opens new possibilities for cost-effective multi-agent systems by dynamically adjusting collaboration topology per query.