Title: FutureWeaver: Strategizing Test-Time Compute for Modularly Collaborative Multi-Agent Systems

Abstract: While increasing test-time computation has proven to boost the performance of large language models (LLMs) without the need for further training, applying these methods to multi-agent systems presents significant hurdles. Current solutions struggle to provide structured mechanisms for distributing computational resources to facilitate effective cooperation, scale coordination efforts, or manage inference costs under strict budgetary limits. To bridge this gap, we introduce FutureWeaver, a framework designed to plan and optimize test-time compute distribution within multi-agent environments subject to fixed budgets. The system leverages collaboration modules—defined as modular, callable functions that bundle reusable multi-agent workflows. These modules are automatically derived through self-play reflection, identifying recurring interaction patterns. Utilizing these modules, FutureWeaver implements a dual-level planning architecture that simultaneously handles short-term action selection and long-term abstract lookahead. This approach optimizes inference trajectories while adhering to budget constraints. Evaluations on complex agent benchmarks reveal that FutureWeaver consistently surpasses baseline methods across various budget configurations, confirming its efficacy in enhancing multi-agent collaboration through inference-time optimization.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC