TUMIX: Multi-Agent Test-Time Scaling

• The paper presents TUMIX’s main contribution: integrating diverse agents (textual, code, and search) that iteratively refine LLM answers through collaborative message passing.
• It introduces an adaptive framework that optimizes accuracy and cost by using LLM-driven agent diversity and a termination criterion based on output confidence and vote margins.
• Empirical results show that TUMIX improves high-difficulty benchmark accuracies by up to 3.55% while significantly reducing inference costs via its multi-agent ensemble approach.

TUMIX (Tool-use Mixture) is a multi-agent test-time scaling framework designed to enhance the reasoning performance of LLMs by orchestrating diverse agent strategies—particularly the mixture of textual, code, and search-based tool use—within a coordinated, iterative ensemble. The central innovation is to run a heterogeneous pool of LLM agents in parallel, each employing distinct tool-use paradigms and answer pathways; these agents recursively share intermediate responses, continually refine their reasoning over multiple rounds, and adaptively halt further refinement based on confidence thresholds. This architecture enables TUMIX to achieve significant accuracy improvements over previous tool-augmented and test-time scaling approaches, while maintaining efficient inference costs (Chen et al., 30 Sep 2025).

1. Multi-Agent, Multi-Tool Framework Architecture

TUMIX structures inference as an ensemble of agents, where each agent is pre-designed to adopt a unique approach to reasoning with LLM tool use. The set of agents typically includes:

• Textual Chain-of-Thought (CoT) agents relying purely on in-context textual reasoning.
• Code Interpreter agents that solve sub-problems via program synthesis and execution.
• Search agents that retrieve and synthesize information from the web or structured databases.
• Hybrid/dual-tool agents that combine, for example, CoT reasoning plus code hints, or chain-of-thought with search augmentation.

Each refinement round, every agent receives:

• The original question,
• Aggregate responses from all agents in the previous round.

After individually generating their updated answers leveraging their designated tool-use “path,” the agents’ responses are pooled for further rounds. The process is terminated based on an adaptive criteria (see Section 5), and the final answer is usually determined via majority voting, though other aggregation heuristics are feasible. This mechanism allows collective exploration of the solution space, cross-pollination of strategies, and robust fallback when one agent class underperforms.

2. Agent Diversity and LLM-Driven Optimization

Agent diversity—defined as the coverage of distinct reasoning modalities and tool application strategies—is fundamental in TUMIX. Unlike repeated sampling from a single best agent, TUMIX employs a curated set of 15 agents spanning pure, hybrid, and tool-specific paradigms. Agent design diversity is further enhanced via LLM-based auto-optimization: LLMs such as Gemini-2.5-Pro are used to prompt and generate additional agent variants tailored to under-explored solution paths. Incorporating these LLM-generated agents empirically yields a further 1.2% accuracy improvement at no substantial additional inference cost.

In formal terms, the TUMIX framework selects the agent ensemble and communication policy π to maximize a utility function balancing accuracy and inference resource consumption: $$\max_\pi~ P\{\text{final answer} = a^*\} - \lambda \cdot \text{Cost}_p$$ where $\lambda$ is a tunable weight and Costₚ models total token consumption and inference count. This formulation allows TUMIX to tailor the agent pool and number of refinement rounds to application-specific cost/accuracy trade-offs.

3. Iterative Answer Refinement and Inter-Agent Communication

The iterative refinement process is realized via a “message passing” protocol. In each round:

• Agents are provided with the cumulative transcript of peer answers.
• Each agent independently processes the transcript and the original question, then produces an updated response using its designated tool-use pipeline.

Agents thus condition on both the evolving multi-agent context and their prior knowledge or reasoning tools, fostering error correction, consensus building, and cross-modal insight propagation. After each round, the output set may be aggregated for answer diversity metrics and used to adapt subsequent round parameters.

A termination mechanism (Section 5) assesses when to stop further refinement rounds—taking into account metrics such as output diversity and voting margins.

4. Empirical Benchmarks and Performance Metrics

TUMIX is validated on three primary, high-difficulty benchmarks:

• HLE (Humanity’s Last Exam): 2,500 multi-domain, challenging questions.
• GPQA (Graduate-Level Google-Proof Q&A): Multiple-choice tasks in fields such as biology, chemistry, and physics.
• AIME: Problems from the American Invitational Mathematics Examination.

Results with Gemini-2.5-Pro show average accuracy improvements of up to 3.55% compared to the strongest tool-augmented and test-time scaling baselines. On HLE, TUMIX increases accuracy from 21.6% (base inference) to 32.3–34.1% (with extended scaling). Accuracy and inference efficiency curves are reported, demonstrating that TUMIX achieves similar or better performance at nearly half the inference cost due to adaptive termination.

Scaling experiments show that extending the number of agents or rounds offers diminishing returns beyond a certain point—a trend also observed in cost-normalized performance analyses.

5. Adaptive Cost Scaling and Termination Criterion

Scalability is managed through an adaptive termination routine, typically instantiated as an LLM-as-Judge: after each refinement round, answer diversity (e.g., vote margin, solution dispersion metrics) and marginal gain (Δ₍ᵣ₎) are measured. If further rounds are unlikely to yield significant improvements, the process halts. This strategy preserves most of the performance improvement at approximately 49% of the full inference cost.

For cases demanding maximum accuracy, a TUMIX+ variant increases the number of inference passes per agent (e.g., varying decoding temperature), further elevating accuracy but raising cost proportionally.

6. Theoretical and Practical Implications

The TUMIX paradigm demonstrates that heterogeneous agent ensembles with diverse tool-use strategies outperform traditional single-strategy or single-agent scaling in reasoning tasks. The explicit formulation:

• Integrates distinct tool-use strategies (text, code, search) into coordinated, parallel pipelines.
• Leverages diversity and LLM-driven agent optimization for robust, cross-domain coverage.
• Exploits adaptive inference scheduling to optimize performance-to-cost ratios.

This modular multi-agent structure is directly applicable to real-world tasks such as research question answering, engineering analysis, or educational assistant systems requiring both accuracy and efficiency. The message-passing and iterative refinement structure generalizes to application-specific domain agents, code assistants, or knowledge synthesis interfaces, where the optimal mixture of reasoning tools is complex and non-obvious.

7. Comparative Position and Future Directions

Relative to extant methods tested in the paper (Majority-Vote, Self-MoA, Symbolic-MoE, DEI, etc.), TUMIX achieves higher and more efficient performance for complex reasoning under comparable compute budgets. The framework’s flexibility in agent design suggests directions for automated agent generation and composition via meta-LLM optimization and for task-adaptive ensemble scaling.

Future advances may explore optimized stopping rules based on dynamic confidence estimation, further integration of verification agents, and domain-specific tool augmentation. TUMIX’s adaptive trade-off framework positions it as a candidate blueprint for agent-centric, resource-aware multi-modal inference in advanced reasoning systems (Chen et al., 30 Sep 2025).