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

Hugging Face Daily Papers Papers
test-time-scaling multi-agent-systems llm-reasoning hierarchical-memory reinforcement-learning inference-compute

Summary

TMAS introduces a multi-agent framework that enhances large language model reasoning by scaling test-time compute through structured collaboration and hierarchical memory systems. The approach uses specialized agents, cross-trajectory information flow, and hybrid reward reinforcement learning to improve iterative scaling and stability on challenging reasoning benchmarks.

Test-time scaling has become an effective paradigm for improving the reasoning ability of large language models by allocating additional computation during inference. Recent structured approaches have further advanced this paradigm by organizing inference across multiple trajectories, refinement rounds, and verification-based feedback. However, existing structured test-time scaling methods either weakly coordinate parallel reasoning trajectories or rely on noisy historical information without explicitly deciding what should be retained and reused, limiting their ability to balance exploration and exploitation. In this work, we propose TMAS, a framework for scaling test-time compute via multi-agent synergy. TMAS organizes inference as a collaborative process among specialized agents, enabling structured information flow across agents, trajectories, and refinement iterations. To support effective cross-trajectory collaboration, TMAS introduces hierarchical memories: the experience bank reuses low-level reliable intermediate conclusions and local feedback, while the guideline bank records previously explored high-level strategies to steer subsequent rollouts away from redundant reasoning patterns. Furthermore, we design a hybrid reward reinforcement learning scheme tailored to TMAS, which jointly preserves basic reasoning capability, enhances experience utilization, and encourages exploration beyond previously attempted solution strategies. Extensive experiments on challenging reasoning benchmarks demonstrate that TMAS achieves stronger iterative scaling than existing test-time scaling baselines, while hybrid reward training further improves scaling effectiveness and stability across iterations. Code and data are available at https://github.com/george-QF/TMAS-code.
Original Article
View Cached Full Text

Cached at: 05/12/26, 10:52 AM

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

Source: https://huggingface.co/papers/2605.10344

Abstract

TMAS is a multi-agent framework for test-time scaling that enhances large language model reasoning through structured collaboration and hierarchical memory systems.

Test-time scalinghas become an effective paradigm for improving thereasoning abilityoflarge language modelsby allocating additional computation during inference. Recent structured approaches have further advanced this paradigm by organizing inference across multiple trajectories, refinement rounds, and verification-based feedback. However, existing structuredtest-time scalingmethods either weakly coordinate parallel reasoning trajectories or rely on noisy historical information without explicitly deciding what should be retained and reused, limiting their ability to balance exploration and exploitation. In this work, we propose TMAS, a framework for scaling test-time compute viamulti-agent synergy. TMAS organizes inference as a collaborative process among specialized agents, enabling structured information flow across agents, trajectories, and refinement iterations. To support effective cross-trajectory collaboration, TMAS introduceshierarchical memories: theexperience bankreuses low-level reliable intermediate conclusions and local feedback, while theguideline bankrecords previously explored high-level strategies to steer subsequent rollouts away from redundant reasoning patterns. Furthermore, we design ahybrid reward reinforcement learningscheme tailored to TMAS, which jointly preserves basic reasoning capability, enhances experience utilization, and encourages exploration beyond previously attempted solution strategies. Extensive experiments on challenging reasoning benchmarks demonstrate that TMAS achieves strongeriterative scalingthan existingtest-time scalingbaselines, while hybrid reward training further improves scaling effectiveness and stability across iterations. Code and data are available at https://github.com/george-QF/TMAS-code.

View arXiv pageView PDFGitHub0Add to collection

Get this paper in your agent:

hf papers read 2605\.10344

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2605.10344 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.10344 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.10344 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Recursive Multi-Agent Systems

Papers with Code Trending

This paper introduces RecursiveMAS, a framework that extends recursive scaling principles to multi-agent systems for improved collaborative reasoning efficiency and accuracy. It demonstrates significant speedups and token reduction across various benchmarks compared to standard baselines.

Multi-Agent Transactive Memory

arXiv cs.AI

Proposes Multi-Agent Transactive Memory (MATM), a framework for population-level storage and retrieval of agent-generated trajectories to improve task performance and reduce interaction steps in interactive environments like ALFWorld and WebArena.

LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling

Hugging Face Daily Papers

This paper introduces AutoTTS, an environment-driven framework that automates the discovery of test-time scaling strategies for LLMs by formulating it as controller synthesis. It demonstrates improved accuracy-cost tradeoffs on mathematical reasoning benchmarks with minimal computational overhead.