← Back to Paper Radar

TUMIX: Multi-Agent Test-Time Scaling with Tool-Use Mixture

Yongchao Chen, Jiefeng Chen, Rui Meng, Jiye Yin, Na Li, Chuchu Fan, Chi Wang, Tomas Pfister, Jinsung Yoon
arXiv.org | 2025
TUMIX is a test-time scaling framework that runs a diverse pool of LLM agents in parallel. Each agent uses a different tool-use strategy (text-only reasoning, code execution, web search, or combinations), and the agents iteratively share and refine answers. Mixing these strategies beats both single tool-augmented agents and standard test-time scaling methods at comparable inference cost.

Problem Statement

Frontier LLMs such as Gemini-2.5-Pro and ChatGPT Agent can call Code Interpreter and Search, but there is little practical guidance on when to reason in text, write code, or search for a given question. Existing test-time scaling methods (e.g., repeated sampling, self-consistency, single-strategy refinement) usually scale one reasoning mode. That leaves out the complementary strengths of different tool-use paths, so compute is used inefficiently and accuracy is capped on diverse reasoning benchmarks.

Key Novelty

  • An ensemble of heterogeneous agents, each with a distinct tool-use strategy and answer path (CoT, code, search, and hybrids), run in parallel with iterative cross-agent sharing and refinement of answers.
  • Evidence that agent diversity and quality drive the gains, plus the use of LLMs to auto-design and optimize the agent pool, which further improves performance.
  • A confidence-based early-stopping mechanism in which an LLM judge halts refinement once consensus or confidence is sufficient, keeping accuracy at about 49% of the inference cost.

Evaluation Highlights

  • Average accuracy improvement of up to 3.55% over the best tool-augmented or test-time scaling baseline on Gemini-2.5-Pro and Gemini-2.5-Flash across key reasoning benchmarks, at near-equal inference cost.
  • Early termination keeps performance at only 49% of the full inference cost; scaling further (more agents or rounds) raises accuracy more but costs more.

Signal Assessment

5/10 This is a solid, practically useful systems contribution that combines multi-agent debate/refinement with tool-use diversity, and it gives clear cost-accuracy evidence on frontier models. Conceptually it is incremental, building on mixture-of-agents, self-consistency, and multi-agent debate, and the gains are moderate (about 3.5%).

Methodology

  1. Build a pool of diverse agents that differ in tool-use strategy: text-only CoT, code-interpreter-based, search-based, and mixed code+search, with varied prompting and answer formats. Optionally expand and optimize the pool with LLM-generated agent designs.
  2. Round 1: run all agents in parallel on the question, each producing an answer through its own tool-use trajectory.
  3. Later rounds: give each agent the original question plus all agents' previous answers so it can refine its response, iterating across rounds.
  4. After each round, use an LLM-as-judge or confidence check to decide whether answers have converged sufficiently, and stop early if so to save compute.
  5. Aggregate final answers (e.g., majority vote across refined agent outputs) to produce the prediction.

System Components

Heterogeneous Agent Pool

A set of agents with distinct tool-use strategies (text reasoning, code execution, web search, hybrids) that provide diverse, complementary reasoning paths.

Iterative Answer Sharing & Refinement

A multi-round loop where each agent sees peers' previous answers and the question, then revises its own answer, so insights pass between tool modalities.

LLM-based Agent Auto-Optimization

Uses an LLM to generate and refine new agent designs (prompts and tool-use patterns), increasing the diversity and quality of the ensemble.

Confidence-based Early Termination

An LLM judge that monitors convergence across rounds and halts refinement when confidence is sufficient, cutting inference cost to about 49% with no loss in accuracy.

Final Answer Aggregation

Combines the refined answers across agents, e.g., by majority voting, into a single output.

Results

Metric/Benchmark Baseline This Paper Delta
Avg. accuracy on key reasoning benchmarks (Gemini-2.5-Pro / Flash) Best tool-augmented / test-time scaling baseline TUMIX Up to +3.55% avg. accuracy
Inference cost (full TUMIX vs. baselines) Comparable compute budget Near-equal cost ≈0% cost overhead
Inference cost with early termination Full TUMIX (100%) ≈49% of full cost ≈51% cost reduction, performance preserved
Further scaling (more agents/rounds) Standard TUMIX TUMIX scaled Higher accuracy at greater cost (qualitative)

Key Takeaways

  • For tool-augmented test-time scaling, diversify how agents reason (text vs. code vs. search) rather than only sampling more from one strategy. Heterogeneous tool-use paths are complementary and give better accuracy per unit of compute.
  • Agent diversity and quality matter more than raw agent count. Using an LLM to propose and refine agent designs is a cheap way to improve the ensemble.
  • Add an LLM-judge-based early-stopping criterion to multi-round refinement pipelines. Many questions converge early, so adaptive termination can roughly halve inference cost without losing accuracy.

Abstract

While integrating tools like Code Interpreter and Search has significantly enhanced Large Language Model (LLM) reasoning in models like ChatGPT Agent and Gemini-Pro, practical guidance on optimal tool use is lacking. The core challenge is effectively combining textual reasoning, coding, and search for diverse questions. In this paper, we propose Tool-Use Mixture (TUMIX), an ensemble framework that runs multiple agents in parallel, each employing distinct tool-use strategies and answer paths. Agents in TUMIX iteratively share and refine responses based on the question and previous answers. In experiments, TUMIX achieves significant gains over state-of-the-art tool-augmented and test-time scaling methods, delivering an average accuracy improvement of up to 3.55% over the best baseline on Gemini-2.5-Pro and Gemini-2.5-Flash across key reasoning benchmarks, with near-equal inference costs. We find that agent diversity and quality are crucial and can be enhanced by using LLMs to auto-optimize agent designs. Furthermore, TUMIX can halt refinement upon reaching sufficient confidence, preserving performance at only 49% of the inference cost. Further scaling can achieve higher performance, albeit at a greater cost.

Generated automatically from available metadata and abstract on 2026-11-09.