| --- |
| license: mit |
| language: |
| - en |
| metrics: |
| - accuracy |
| base_model: |
| - Qwen/Qwen2.5-3B-Instruct |
| pipeline_tag: reinforcement-learning |
| tags: |
| - agent |
| - reinforcement-learning |
| - routing |
| - tool-use |
| - reasoning |
| --- |
| |
| # ATLAS-RL: Reinforcement Learning for Adaptive Model–Tool Routing |
|
|
| This model is trained using the **ATLAS** framework proposed in the paper: |
|
|
| **ATLAS: Adaptive Tool–LLM Alignment and Synergistic Invocation via Reinforcement Learning** |
|
|
| 📄 **Paper:** https://huggingface.co/papers/2601.03872 |
| 📄 **arXiv:** https://arxiv.org/abs/2601.03872 |
|
|
| --- |
|
|
| ## Overview |
|
|
| ATLAS-RL is a reinforcement learning–trained routing model designed to coordinate large language models and external tools for complex reasoning tasks. |
|
|
| Instead of relying on static tool selection or predefined workflows, ATLAS learns a **policy that dynamically decides which model or tool to invoke** during multi-step reasoning. The policy is trained with reinforcement learning to optimize final task accuracy while maintaining efficient routing behavior. |
|
|
| Through RL optimization, the model learns to: |
|
|
| - select appropriate models for different problem types |
| - decide when external tools should be invoked |
| - coordinate multi-step reasoning across heterogeneous components |
|
|
| This enables more effective collaboration between language models and tools in tasks such as mathematical reasoning, coding, and scientific question answering. |
|
|
| --- |
|
|
| ## Experimental Results |
|
|
| The following tables summarize the performance of ATLAS compared to baseline routers. ATLAS (RL) demonstrates superior generalization, especially in Out-of-Distribution (OOD) scenarios. |
|
|
| ### In-Distribution Performance |
| | Model | avg score | |
| | :--- | :--- | |
| | RouterDC | 53.4 | |
| | MLPRouter | 42.3 | |
| | BertRouter | 48.6 | |
| | **ATLAS (cluster)** | **63.5** | |
|
|
| ### Out-of-Distribution Performance |
| | Model | avg score | |
| | :--- | :--- | |
| | RouterDC | 46.3 | |
| | MLPRouter | 39.3 | |
| | BertRouter | 43.6 | |
| | ATLAS (cluster) | 49.2 | |
| | **ATLAS (RL)** | **59.4** | |
|
|
| --- |
|
|
| ## Key Features |
|
|
| - 🔁 **RL-based Routing Policy** Learns routing strategies through reinforcement learning rather than heuristic rules. |
|
|
| - 🧠 **Adaptive Model–Tool Coordination** Dynamically determines which models or tools should be used during reasoning. |
|
|
| - ⚡ **Multi-step Decision Making** Supports iterative reasoning where multiple routing decisions may occur within a single query. |
|
|
| - 🌍 **Generalizable Routing Strategies** The learned policy transfers to unseen tasks without manual routing design. |
|
|
| --- |
|
|
| ## Reinforcement Learning Framework |
|
|
| ATLAS trains a routing policy using reinforcement learning to optimize decision making during reasoning. |
|
|
| At each step, the policy selects the next action (e.g., invoking a model or tool). The policy is optimized using reward signals that capture answer correctness, structured outputs, and routing efficiency. |
|
|
| The training objective encourages the model to discover **effective collaboration patterns between models and tools**, improving reasoning performance across diverse tasks. |
|
|
|
|
|
|
| --- |
|
|
| ## Citation |
|
|
| If you use this model or the ATLAS framework in your research, please cite: |
|
|
| ```bibtex |
| @article{wu2026atlas, |
| title={Atlas: Orchestrating Heterogeneous Models and Tools for Multi-Domain Complex Reasoning}, |
| author={Wu, Jinyang and Zhai, Guocheng and Jin, Ruihan and Yuan, Jiahao and Shen, Yuhao and Zhang, Shuai and Wen, Zhengqi and Tao, Jianhua}, |
| journal={arXiv preprint arXiv:2601.03872}, |
| year={2026} |
| } |
| |
| ``` |
