TMAS: Scaling Test-Time Compute via Multi-Agent Synergy

Researchers introduce TMAS, a multi-agent framework that improves test-time compute scaling for large language models by enabling specialized agents to collaborate through hierarchical memory systems. The approach balances exploration and exploitation more effectively than existing methods, achieving stronger iterative scaling on challenging reasoning benchmarks.

TMAS addresses a fundamental challenge in modern AI: how to optimally allocate computation during inference to improve reasoning capabilities. Traditional test-time scaling approaches either fail to coordinate multiple reasoning paths effectively or rely on poorly managed historical information, creating inefficiencies in the search for correct solutions. This research demonstrates that organizing inference as a collaborative multi-agent process with structured information flow can meaningfully improve performance.

The framework's innovation lies in its dual-memory architecture. The experience bank preserves verified intermediate results and local feedback, while the guideline bank tracks high-level strategies, preventing redundant exploration. This design mirrors human problem-solving, where experts retain both specific facts and general heuristics. The hybrid reward reinforcement learning scheme further optimizes the system by simultaneously preserving base reasoning ability, improving experience reuse, and encouraging novel solution approaches.

For the AI development community, TMAS represents progress toward more efficient inference-time reasoning without requiring larger base models. This has practical implications for cost reduction and deployment scalability, particularly for applications requiring complex multi-step reasoning like mathematical problem-solving and logic tasks. As language models become more prevalent in production systems, test-time compute optimization directly impacts operational costs and latency constraints.

The open-source release of code and data accelerates reproducibility and adoption. Future development likely focuses on applying this framework to broader reasoning domains and combining it with other inference optimization techniques. Whether TMAS substantially influences production AI systems depends on how well it generalizes beyond benchmark tasks and integrates with existing LLM deployment pipelines.

• →TMAS framework enables multi-agent collaboration during LLM inference with hierarchical memory systems for improved reasoning
• →Dual-memory architecture (experience bank and guideline bank) prevents redundant computation while preserving useful intermediate results
• →Hybrid reward training balances exploration of new strategies with exploitation of known solutions for stronger iterative scaling
• →Test-time compute scaling offers cost-effective performance improvements without requiring larger base models
• →Open-source release enables rapid adoption and validation across diverse reasoning tasks