---
title: 'TUMIX: Multi-Agent Tool-Use Framework'
url: https://www.emergentmind.com/topics/tool-use-mixture-tumix
type: topic
---

# TUMIX: Multi-Agent Tool-Use Framework

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 large language model (LLM) platforms such as Gemini-2.5-Pro and Gemini-2.5-Flash [2510.01279].

## 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 ($\operatorname{Cov}(Z_i, Z_j) > 0$).

The TUMIX policy, under total inference-cost budget $C$, determines:
- Agent invocation schedule per round,
- Inter-agent communication graph,
- Dynamic halting criterion for total rounds $R$,
- Aggregation rule for final answer selection $\hat{a}$.

The canonical objective is to maximize net utility:
$$
\max_\pi \; P\{\hat{a} = a^*\} - \lambda \cdot \text{Cost}_\pi
$$
where $\text{Cost}_\pi = \sum_{r=1}^R \sum_{i \in S_r} c_i$, $c_i$ denotes the per-agent inference cost in tokens/API calls, and $S_r \subseteq A$ is the active agent set at round $r$. In typical instantiations, all agents participate in all rounds ($S_r = A$, $\forall r$), yielding nearly constant per-round cost and majority-vote aggregation:
$$
\hat{a} = \arg\max_{y \in \mathcal{A}} \sum_{i \in A} 1\{y_i = y\}
$$
The framework admits confidence- or cost-weighted fusion, but empirical results indicate majority vote is robust within this paradigm [2510.01279].

## 2. Iterative Refinement Algorithm

TUMIX operates via synchronous, parallel rounds. In each round $r$, every agent $A_i$ receives both the original question $q$ and the full set of prior round answers $\mathcal{R}_{r-1} = \{y^{1}_{r-1}, \ldots, y^K_{r-1}\}$. Agents then output $y^i_r = A_i(q \parallel \mathcal{R}_{r-1})$. This process embodies message-passing with tool-augmented reasoning up to a configurable tool budget $R_{\text{tool}}$.

### Pseudocode Overview (answers-only view):
```plaintext
Inputs: question q, agent set A={A_i}_i=1^K
Hyperparams: r_min=2, r_max; tool-budget R_tool; timeout τ
ℛ₀←∅
for r=1…r_max:
  in parallel ∀i∈A:
    prompt_i ← BuildPrompt(q, ℛ_{r−1})
    y_r^i ← AgentCall(A_i, prompt_i, R_tool, τ)
  ℛ_r ← {y_r^i : i∈A}
  if r ≥ r_min and LLMTerminate(q,ℛ_r)=STOP: break
output  \hat a = MajorityVote(ℛ_r)
```
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 ($S_0$), the LLM is tasked to generate $M = 25$ additional, structurally diverse candidates, yielding a pool of 40 ($S_0 \cup S_{\text{LLM}}$). These are evaluated on first-round accuracy and coverage using the HLE benchmark. The top 15 are selected by a combined metric:
$$
\text{CombinedScore}_j = \frac{\text{Coverage}_j}{E[\text{Coverage}]} + \frac{\text{Acc}_j}{E[\text{Acc}]}
$$
where $\text{Coverage}_j = P(\cup_i \{y^j_i = a^*\})$. 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 $2$–$3$ 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 $r$ is:
$$
\Delta_r = E[A_{r+1} - A_r | \text{signals up to } r]
$$
Halting is triggered when $\Delta_r \leq \lambda \cdot \text{(cost per round)}$. Operationally, after $r \geq r_{\min} = 2$ rounds, an “LLM-as-Judge” prompt assesses whether all $K$ answers are in strong consensus; further refinement occurs only if differences persist. This stopping strategy achieves essentially maximal accuracy with only $49\%$ of the inference calls and $46\%$ of the token budget relative to non-adaptive $5$-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 $+3.55\%$ gain on HLE at matched cost over the best baseline ($29.5\% \rightarrow 32.3\%$). On Gemini-2.5-Flash, relative improvements are even greater ($+6.9\%$). Ablation studies affirm that agent diversity is indispensable (shrinking to three or one agent reduces accuracy by $2$–$3\%$ absolute), and that adaptive halting cuts cost by more than half with no loss in accuracy. The evolved agent pool confers a further $\approx 1.2\%$ lift. Scaling agent ensemble size in TUMIX+ achieves peak accuracy ($34.1\%$ 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 3), illustrating sharp convergence after round $2$.
- Sankey diagram (Figure 4), tracking agent consensus and correctness across refinement stages.
- Scaling law (Figure 7), 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 [2510.01279].

Source: https://www.emergentmind.com/topics/tool-use-mixture-tumix