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

TL;DR

Atlas addresses the challenge of dynamically orchestrating heterogeneous LLMs and external tools for multi-domain reasoning by combining a training-free cluster-based routing path with an RL-driven multi-step routing path. The dual-path design leverages domain priors and empirical synergies while enabling open-domain generalization through learned policies, achieving state-of-the-art or competitive performance across 15 benchmarks and robust multimodal reasoning. Key contributions include explicit joint optimization of model-tool pairs, a composite reward scheme that decouples routing efficiency from task correctness, and demonstration of strong generalization to expanded pools without retraining. The results highlight the practical potential of ecosystem-centric orchestration to harness heterogeneous capabilities for efficient, scalable autonomous reasoning.

Abstract

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.

Figures (14)

• Figure 1: Comparison of different LLM inference paradigms. While routing (efficiency) and RL (performance optimization) present a promising approach, dynamic tool usage still faces significant challenges.
• Figure 2: Performance comparison on in-distribution and out-of-distribution settings. Our Atlas method consistently outperforms all baselines across diverse datasets, demonstrating superior generalization capability.
• Figure 3: Overview of Adaptive Tool-LLM Alignment and Synergistic Invocation (Atlas). The framework operates via a dual-path approach: (1) Training-free Cluster-based Routing; and (2) RL-driven Multi-step Routing.
• Figure 4: Performance comparison of Atlas against single-tool baselines across multi-modal benchmarks. 'None' denotes direct reasoning without any tools. Atlas achieves the highest accuracy.
• Figure 5: Analysis of LLM API call count and Atlas(RL) training dynamics.