PaperlayerView exact revision on arXiv

Source-linked AI summary

Atlas: Orchestrating Heterogeneous Models and Tools for Multi-Domain Complex Reasoning

Jinyang Wu, Guocheng Zhai, Ruihan Jin, Jiahao Yuan, Yuhao Shen, Shuai Zhang, Zhengqi Wen, Jianhua Tao

arXiv:2601.03872v2cs.CL

TL;DR

Existing methods struggle to select effective combinations across increasingly diverse LLMs and tools, often treating models or tool use in isolation. ATLAS addresses this with training-free cluster-based routing for domain-specific alignment and RL-driven multi-step routing for open-domain exploration. Across multi-domain benchmarks, it outperforms strong baselines and closed-source models, while its scope remains limited by modality coverage and reliance on reliable API access.

  • Problem

    Existing approaches overlook the synergies among heterogeneous LLMs and tools, making optimal model-tool selection a high-dimensional reasoning challenge.

  • Method

    ATLAS uses training-free cluster-based routing for domain-specific selection and RL-driven multi-step routing for unfamiliar tasks requiring iterative exploration.

  • Results

    Across 9 tasks and 15 benchmarks, ATLAS outperforms top-performing closed-source LLMs and routing methods and adapts robustly to multi-modal scenarios.

  • Takeaways & Limitations

    Strategic coordination of heterogeneous model-tool combinations can unlock superior reasoning while maintaining efficiency across diverse benchmarks.

  • Takeaways & Limitations

    Evaluation focuses primarily on text-based and visual reasoning, while real-world performance assumes reliable API access and may be affected by latency or service unavailability.

Abstract

from arXiv · show

The integration of large language models (LLMs) with external tools has significantly expanded the capabilities of AI agents. However, as the diversity of both LLMs and tools increases, selecting the optimal model-tool combination becomes a high-dimensional optimization challenge. Existing approaches often rely on a single model or fixed tool-calling logic, failing to exploit the performance variations across heterogeneous model-tool pairs. In this paper, we present ATLAS (Adaptive Tool-LLM Alignment and Synergistic Invocation), a dual-path framework for dynamic tool usage in cross-domain complex reasoning. ATLAS operates via a dual-path approach: (1) \textbf{training-free cluster-based routing} that exploits empirical priors for domain-specific alignment, and (2) \textbf{RL-based multi-step routing} that explores autonomous trajectories for out-of-distribution generalization. Extensive experiments across 15 benchmarks demonstrate that our method outperforms closed-source models like GPT-4o, surpassing existing routing methods on both in-distribution (+10.1%) and out-of-distribution (+13.1%) tasks. Furthermore, our framework shows significant gains in visual reasoning by orchestrating specialized multi-modal tools.

1 Introduction

ATLAS addresses the difficulty of selecting effective combinations of heterogeneous LLMs and tools by combining training-free, domain-specific routing with RL-driven multi-step exploration. Across diverse tasks and benchmarks, it is reported to outperform strong baselines and closed-source LLMs.

  • Motivation: Existing routing methods select models without integrating external tools, while tool-use frameworks rely on fixed invocation logic that cannot adapt model-tool combinations dynamically.These limitations become more consequential as the diversity of LLMs and tools increases.
  • Approach: Training-free cluster-based routing uses domain-specific expertise and historical performance patterns to select model-tool pairs efficiently in familiar domains.The method operates in a semantic embedding space and leverages empirical priors for rapid routing.
  • Approach: RL-driven multi-step routing iteratively explores model-tool interactions for generalized reasoning when explicit domain priors are absent.This path targets open-domain and unfamiliar tasks through iterative exploration.
  • Approach: ATLAS dynamically orchestrates model-tool combinations through a generalizable framework for complex reasoning tasks.The framework explicitly optimizes heterogeneous synergies and supports adaptive tool invocation.
  • Results: Across 9 tasks and 15 benchmarks, ATLAS outperforms top-performing closed-source LLMs and routing methods while adapting robustly to multi-modal scenarios.The reported evaluation includes both in-distribution and out-of-distribution settings.

2 Related Work

Prior work separately improves query-based model routing, tool use, and reinforcement-learning-based reasoning, but these approaches commonly overlook joint model-tool synergies. ATLAS extends this line by jointly optimizing heterogeneous model-tool combinations through dual-path routing.

  • Query-based LLM Routing: Query-based routing balances performance and computational efficiency through reward-guided, contrastive, cascading, adaptive, and budget-allocation strategies.These methods dynamically select or allocate models at query level but often treat LLMs as isolated execution units.
  • ATLAS: ATLAS combines training-free cluster-based routing for quick decisions with RL-driven multi-step routing for complex open-domain tasks requiring iterative interactions.The two-tier strategy connects efficient routing with broader exploration.
  • Query-based LLM Routing: Existing routing methods neglect synergies between specific model capabilities and external tools.Their model-centric treatment limits integration with tool capabilities.
  • Reinforcement Learning for LLM: Reinforcement learning has progressed from reward-model-based methods such as RLHF and PPO to preference-learning approaches such as DPO for optimizing LLM behavior and reasoning.The cited approaches streamline or improve preference and reasoning optimization in different ways.
  • Reinforcement Learning for LLM: Prior RL methods primarily optimize single-model or single-tool settings, overlooking combined model-tool synergies.ATLAS extends RL to jointly optimize model-tool combinations for more adaptive and efficient reasoning.

3 Methodology

ATLAS combines training-free cluster routing for quick model–tool selection with RL-driven multi-step routing for complex tasks requiring iterative interactions. Its routing uses semantic structure, empirical performance and cost signals, while the RL path learns autonomous orchestration through composite rewards.

  • 3 Methodology: ATLAS uses a two-tier strategy: cluster-based routing for quick decisions and RL-driven multi-step routing for complex iterative model-tool interactions.The cluster path targets low-latency orchestration, while the RL path addresses tasks requiring multiple reasoning and tool-use rounds.
  • 3.1 Training-Free Cluster-Based Routing: The search space is the Cartesian product S = M × T of candidate LLMs and available tools, reflecting query-dependent model-tool routing.ATLAS approximates the routing function by exploiting semantic locality and historical metadata.
  • 3.1 Training-Free Cluster-Based Routing: Training queries are embedded into a latent manifold and partitioned into clusters whose centroids group similar reasoning requirements and tool affinities.The method derives empirical statistics for each model-tool pair within each cluster.
  • 3.1 Training-Free Cluster-Based Routing: Cluster utility balances empirical accuracy against operational cost using α, with costs computed from average input/output token throughput and unit prices.The utility score supports a flexible performance-cost trade-off during pair selection.
  • 3.1 Training-Free Cluster-Based Routing: At inference, ATLAS embeds a query, assigns it to the nearest cluster, and retrieves the optimal cached model-tool pair with constant-time complexity relative to cluster count.Caching heterogeneous synergies enables real-time, cost-aware tool invocation.
  • 3.2 RL-Driven Multi-Step Routing: The RL router treats orchestration as sequential decision-making and trains with PPO using format, outcome and model-selection rewards.These signals enforce structured execution, task correctness and efficient model choice, while routing remains autonomous across internal reasoning and external invocation.

4 Experiments

ATLAS is evaluated across multi-domain, multimodal, routing-pool, reasoning-boundary, efficiency, and statistical analyses. Results show strong in-distribution performance, improved OOD transfer, gains from specialized tools and pool extensions, efficient exploration, adaptive invocation, and statistically significant improvements.

  • Evaluation setup: ATLAS is evaluated across diverse text and visual benchmarks using heterogeneous open-source models and foundation or multimodal tools.The evaluation covers 15 benchmarks, including mathematical, coding, arithmetic, commonsense, logical, scientific, chart, geometry, counting, and table-understanding tasks.
  • In-distribution performance: 63.5% average accuracy lets ATLAS(cluster) surpass RouterDC by 10.1% in-distribution.ATLAS reaches 40.0% on AIME25 and 82.5% on AMC, exceeding RouterDC by 16.7% and 20.0%, respectively.
  • Generalization scenarios: 59.4% average accuracy lets ATLAS(RL) exceed ATLAS(cluster) by 10.2% and RouterDC by 13.1% out-of-distribution.On AIME24 and AIME25, ATLAS(RL) achieves 43.3% and 33.3%, compared with 13.3% and 3.3% for clustering.
  • Multimodal orchestration: 68.9% average accuracy lets ATLAS outperform the strongest single-tool baseline by 4.3% across five visual reasoning datasets.ATLAS exceeds individual tools across task categories, including Qwen3-Chart on ChartQA and Geometry3K.
  • Routing-pool extensions: 61.7% accuracy after pool extension represents a 2.3% gain for ATLAS(RL), with AIME24 and AIME25 each improving by 6.7%.The added math-specialized model and verification tool improve performance, while RandomRouter degrades and BertRouter gains only 1.5%.
  • Reasoning capacity: +23.0% absolute pass@1 improvement, from 36.4% to 59.4%, demonstrates more efficient exploration.At pass@16, the upper bound reaches 63.1%, while task-level gains range from +6.1% to +35.0%.
  • Efficiency and training dynamics: ATLAS increases API calls for challenging reasoning tasks but keeps calls minimal for straightforward retrieval tasks.Reward shaping accelerates convergence and entropy reduction, moving routing toward higher-confidence decisions.
  • Statistical significance: ATLAS’s superiority over RouterDC is statistically significant in both in-distribution and OOD settings.The in-distribution Wilcoxon signed-rank test reports p = 9.7 × 10^-4, below α=0.05.

5 Conclusion

ATLAS introduces a dual-path framework for dynamic model-tool alignment: cluster-based routing provides domain-specific efficiency, while RL-driven routing supports open-domain adaptability. Across diverse benchmarks, strategic coordination of heterogeneous models and tools rivals or exceeds powerful closed-source systems while maintaining efficiency.

  • 5 Conclusion: ATLAS combines training-free cluster-based routing for domain-specific efficiency with RL-driven exploration for open-domain adaptability.The framework dynamically aligns heterogeneous models and tools rather than relying on a single model or fixed invocation logic.
  • 5 Conclusion: Strategic coordination of heterogeneous model-tool combinations unlocks superior reasoning while maintaining efficiency.The conclusion frames ATLAS as a shift from model-centric scaling toward ecosystem-centric orchestration.

Limitations

The evaluation focuses on text-based and visual reasoning, while other modalities remain unexplored. Real-world performance also assumes reliable API access, making latency and service unavailability relevant deployment constraints.

  • Scope: ATLAS is evaluated primarily on text-based and visual reasoning, leaving audio and video unexplored.The authors identify extension to additional modalities as future work.
  • Deployment assumptions: Reliable API access is assumed, so network latency or service unavailability may affect real-world performance.The authors propose investigating lightweight policies and robust fallback mechanisms.

Ethical Considerations

ATLAS leaves the underlying candidate models and tools unmodified, so it inherits their biases, safety limitations, and ethical concerns. The paper recommends evaluating candidates for compliance and applying safety measures in deployment.

  • ATLAS comprises a training-free cluster-based router and an RL-trained policy model that orchestrates existing LLMs and tools.
  • Because the candidate models and tools remain unmodified, ATLAS inherits their potential biases, safety limitations, and ethical concerns.
  • The paper recommends evaluating candidate components for ethical compliance and applying appropriate safety measures in real-world deployment.

A.2 Detailed Specification of Reward Signals

ATLAS trains its routing policy with composite signals for format compliance, task correctness, and model-selection efficiency. The rewards constrain trajectory structure, evaluate outcomes, and penalize suboptimal model choices.

  • The composite reward is rϕ = Rfmt + γRout + ξRsel, combining format, outcome, and model-selection signals.The formulation bridges structured interaction with task-specific accuracy.
  • Rout is a binary trajectory-level signal that indicates whether the completed answer is correct.
  • Rsel penalizes suboptimal model invocations relative to a task-specific optimal model.MBPP uses Qwen2.5-Coder-7B-Instruct, while Calculator and NQ identify the best candidate through offline GPT-4o evaluation.

B.1 Datasets

The evaluation covers diverse programming, mathematics, knowledge, logic, multimodal, and tool-use benchmarks. ATLAS routes queries through semantic clusters and model-tool execution, with experiments examining scaling, multimodal orchestration, and out-of-distribution behavior.

  • Datasets: AIME 2024 and AIME 2025 each contain 30 problems requiring advanced problem-solving, strategic thinking, and precise numerical computation.
  • Routing procedure: The cluster-based router embeds each query, assigns it to the nearest semantic cluster, selects a model-tool pair, and invokes it.
  • Datasets: The benchmark suite spans programming, mathematics, knowledge retrieval, logical reasoning, visual question answering, and arithmetic computation.Examples include HumanEval, MBPP, AIME, NQ, LogiQA2, ChartQA, Geometry3K, TallyQA, CountBench, TableVQA, and Calculator.
  • Routing procedure: The RL routing algorithm performs multi-step reasoning with dynamic model-tool invocation and parses the final answer from the trajectory.
  • Evaluation: Across multimodal evaluations, ATLAS consistently outperforms single-tool baselines under matched Qwen3-8B-VL backbones and identical protocols.The comparison is designed to attribute gains to adaptive tool orchestration rather than model-level differences or prompt-engineering artifacts.
  • Evaluation: Self-consistency performance generally increases with sample count, including an AIME24 improvement from 43.3% to 70.0% at SC@16.

C.4 Analysis of Model-Tool Alignment Preferences

ATLAS aligns specialized model-tool pairs with deterministic domains while adapting its selections for knowledge-intensive and broad-spectrum tasks. Its RL policy also scales to larger tool pools, and reward ablations identify format enforcement as especially important.

  • Model-tool preferences: In coding and advanced mathematics, ATLAS selects specialized pairings such as Qwen2.5-Coder-7B with Python or DeepSeek-R1 with PRM in nearly 100% of cases.
  • Model-tool preferences: Knowledge-intensive tasks favor retrieval-augmented configurations, primarily pairing Llama-3.1-8B with Web-Search.
  • Model-tool preferences: For LQA2, the dominant ATLAS(RL) combination accounts for 56.4%, indicating more distributed selections on broad-spectrum tasks.
  • Tool-pool scaling: +1.5% average accuracy accompanies growth from 4 to 8 tools, while latency increases by only ∼2 s.The reported scaling result indicates sub-linear latency growth as the action space expands.
  • Tool-calling evaluation: 86.0% is ATLAS(RL)'s average BFCL single-round score, surpassing all listed closed-source models.
  • Reward ablations: Removing Rsel reduces performance by 3.1% from 59.4% to 56.3%, while removing Rfmt causes a larger 6.1% drop from 59.4% to 53.3%.The Rfmt ablation also reports drops on AIME24, AMC, and HumanEval; the paper characterizes Rsel as optional efficiency guidance.
  • Robustness analyses: Cluster-based routing achieves its best reported average accuracy at K = 8 with 63.5%, compared with 51.9% at K = 4 and 62.8% at K = 16.
  • Robustness analyses: Embedding-model performance varies minimally from 63.1% to 63.9%, and Wilcoxon tests report significant differences against ROUTERDC in both in-domain and OOD settings.

D Case Study

ATLAS dynamically orchestrates model-tool combinations across reasoning tasks, including iterative self-correction and task-specific specialization. The case studies illustrate adaptive routing for verification, computation, retrieval, code generation, and mathematical reasoning.

  • Task-Aware Model-Tool Alignment and Selection: The examples collectively indicate that ATLAS associates task categories with suitable model-tool configurations rather than relying on rigid invocation rules.
  • Adaptive Multi-turn Reasoning: ATLAS recovers from contradictory feedback by switching from Qwen2.5-7B with web search to InternLM3-8B, ultimately selecting the correct answer.The logical-reasoning example demonstrates multi-turn deliberation and self-correction after an initially suboptimal hypothesis.
  • Task-Aware Model-Tool Alignment and Selection: For arithmetic computation, ATLAS directly invokes the calculator tool without unnecessary reasoning steps.
  • Task-Aware Model-Tool Alignment and Selection: For factual retrieval, ATLAS routes to Llama-3.1-8B with web search to obtain external knowledge.
  • Task-Aware Model-Tool Alignment and Selection: For code generation, ATLAS delegates to Qwen2.5-Coder with Python execution, while challenging mathematical problems use DeepSeek-7B with PRM verification.

E.1 Distinguishing ATLAS from Prior Routing and Tool Usage Methods

ATLAS is presented as a joint model-tool routing framework that combines cluster-based routing with reinforcement-learning exploration. Its design addresses prior methods' separate treatment of models and tools while retaining routing-efficiency signals and identifying several future evaluation boundaries.

  • Distinguishing ATLAS from Prior Routing and Tool Usage Methods: ATLAS jointly optimizes heterogeneous model-tool combinations, unlike prior routing methods that select models in isolation and tool frameworks that use fixed invocation logic.
  • Distinguishing ATLAS from Prior Routing and Tool Usage Methods: Its dual-path architecture combines training-free cluster-based routing for domain-specific priors with RL-driven exploration for unfamiliar tasks.
  • Distinguishing ATLAS from Prior Routing and Tool Usage Methods: The composite reward rϕ = Rfmt + γRout + ξRsel balances structured execution, task correctness, and routing efficiency.
  • Distinguishing ATLAS from Prior Routing and Tool Usage Methods: Rsel is reported to accelerate convergence and reduce entropy, while expanded model-pool results indicate transferable routing principles rather than memorized mappings.
  • Distinguishing ATLAS from Prior Routing and Tool Usage Methods: Cluster-based routing achieves 63.5% average accuracy with zero training cost when domain-specific training data is available.
  • Distinguishing ATLAS from Prior Routing and Tool Usage Methods: Future evaluation should address retrieval noise, RL sample efficiency, long-horizon interactive settings, inference latency, and richer perceptual abilities.
Loading 2601.03872v2…