TUMIX: Multi-Agent Tool-Use Framework

• TUMIX is a framework that ensembles multiple agent-tool strategies, integrating chain-of-thought, code interpreters, and web search for advanced reasoning.
• It operates in synchronous, parallel rounds with iterative answer sharing and adaptive halting to efficiently balance accuracy and cost.
• Empirical results show TUMIX improves accuracy by up to 6.9% and reduces inference cost by over 50% compared to traditional multi-agent methods.

Tool-Use Mixture (TUMIX) is an ensemble framework designed for multi-agent test-time scaling with tool augmentation, achieving accurate and cost-effective reasoning through parallelized and iteratively refined agent collaboration. It integrates disparate tool-use strategies—including text-only chain-of-thought (CoT), code interpreters, and web search—within a multi-agent protocol, leveraging iterative answer sharing and automated agent design to optimize both accuracy and efficiency for LLM platforms such as Gemini-2.5-Pro and Gemini-2.5-Flash (Chen et al., 30 Sep 2025).

1. Formal Framework and Objective

TUMIX formalizes multi-agent tool use as a policy $\pi$ operating over a fixed pool of $K$ agents $A = \{A_1, \ldots, A_K\}$, each employing a unique tool-use paradigm. For a given question $q$ with true answer $a^*$, agent $A_i$ produces a candidate response $y_i \in \mathcal{A}$ according to its stochastic output distribution $P_i(y \mid q)$. The agent's latent competence is defined by $p_i(q) = P_i(y_i = a^*)$, and answer correctness is captured by $Z_i = 1\{y_i = a^*\}$. Due to shared LLM backbones, pairwise agent correlations are nontrivial ($K$0).

The TUMIX policy, under total inference-cost budget $K$1, determines:

• Agent invocation schedule per round,
• Inter-agent communication graph,
• Dynamic halting criterion for total rounds $K$2,
• Aggregation rule for final answer selection $K$3.

The canonical objective is to maximize net utility:

$K$4

where $K$5, $K$6 denotes the per-agent inference cost in tokens/API calls, and $K$7 is the active agent set at round $K$8. In typical instantiations, all agents participate in all rounds ($K$9, $A = \{A_1, \ldots, A_K\}$0), yielding nearly constant per-round cost and majority-vote aggregation:

$A = \{A_1, \ldots, A_K\}$1

The framework admits confidence- or cost-weighted fusion, but empirical results indicate majority vote is robust within this paradigm (Chen et al., 30 Sep 2025).

2. Iterative Refinement Algorithm

TUMIX operates via synchronous, parallel rounds. In each round $A = \{A_1, \ldots, A_K\}$2, every agent $A = \{A_1, \ldots, A_K\}$3 receives both the original question $A = \{A_1, \ldots, A_K\}$4 and the full set of prior round answers $A = \{A_1, \ldots, A_K\}$5. Agents then output $A = \{A_1, \ldots, A_K\}$6. This process embodies message-passing with tool-augmented reasoning up to a configurable tool budget $A = \{A_1, \ldots, A_K\}$7.

Pseudocode Overview (answers-only view):

$A_i$1 The AgentCall operation supports multimodal tool use: execution of code blocks (≤60s), search-query expansion, and agent-internal iterative self-reflection, followed by mandatory answer emission upon budget exhaustion.

3. Agent Diversity and LLM-Driven Design Optimization

Rather than exclusively relying on manually crafted agent configurations, TUMIX leverages LLM-driven proposal mechanisms for agent design. Starting from 15 human-designed agents ($A = \{A_1, \ldots, A_K\}$8), the LLM is tasked to generate $A = \{A_1, \ldots, A_K\}$9 additional, structurally diverse candidates, yielding a pool of 40 ($q$0). These are evaluated on first-round accuracy and coverage using the HLE benchmark. The top 15 are selected by a combined metric:

$q$1

where $q$2. This greedy, non-gradient selection process defines a static set of "evolved" agents ("Evolved set", Editor's term), superseding dynamic agent sampling strategies, which empirically offer limited further gains.

4. Dynamic Halting and Refinement Stopping

Empirical analysis reveals that after $q$3–$q$4 refinement rounds, agent answer diversity contracts rapidly—agents converge, correctly or otherwise, to a consensus. To economize inference expenditure, an adaptive halting rule is invoked. The expected marginal gain for round $q$5 is:

$q$6

Halting is triggered when $q$7. Operationally, after $q$8 rounds, an “LLM-as-Judge” prompt assesses whether all $q$9 answers are in strong consensus; further refinement occurs only if differences persist. This stopping strategy achieves essentially maximal accuracy with only $a^*$0 of the inference calls and $a^*$1 of the token budget relative to non-adaptive $a^*$2-round protocols.

5. Empirical Results and Ablation Analysis

Experiments were conducted using Gemini-2.5-Pro and Gemini-2.5-Flash, both with full code interpreter and Google Search integration. Benchmarks include:

• HLE (2,500 broad-domain “Last Exam” questions)
• GPQA-Diamond (198 expert-curated multiple-choice questions in STEM)
• AIME 24–25 (60 advanced mathematics problems)

Comparison against baselines—Majority-Vote, GSA, Self-Reflection, SETS, Self-MoA, Symbolic-MoE, DEI, and SciMaster—at cost-matched levels yields:

Method	HLE	GPQA	AIME	AvgNorm
w/o TTS	21.6%	84.6%	87.3%	64.5
Best Baseline	29.5%*	86.9%	95.0%	70.3
TUMIX	32.3%	87.9%	96.7%	72.3
TUMIX-Evolve	32.7%*	88.1%	96.7%	72.5
TUMIX+ (scaled)	34.1%	88.3%	96.7%	73.0

TUMIX delivers a $a^*$3 gain on HLE at matched cost over the best baseline ($a^*$4). On Gemini-2.5-Flash, relative improvements are even greater ($a^*$5). Ablation studies affirm that agent diversity is indispensable (shrinking to three or one agent reduces accuracy by $a^*$6–$a^*$7 absolute), and that adaptive halting cuts cost by more than half with no loss in accuracy. The evolved agent pool confers a further $a^*$8 lift. Scaling agent ensemble size in TUMIX+ achieves peak accuracy ($a^*$9 on HLE), but at more than quadruple inference cost.

6. Analytical Visualizations and Scaling Properties

Key empirical analyses include:

• Bar charts (Figure 1) comparing equal-cost accuracy of all methods on Gemini-Pro and Gemini-Flash across benchmarks.
• Coverage and average accuracy per round (Figure 2), illustrating sharp convergence after round $A_i$0.
• Sankey diagram (Figure 3), tracking agent consensus and correctness across refinement stages.
• Scaling law (Figure 4), relating accuracy improvements to inference calls and token count, establishing TUMIX’s efficiency over prior test-time scaling approaches.

Collectively, these results substantiate the superiority of heterogeneous, parallel tool-use with iterative refinement and adaptive halting, validating the TUMIX architecture as an optimal blend of accuracy and cost-efficiency for multi-tool, multi-agent LLM reasoning (Chen et al., 30 Sep 2025).