File size: 8,863 Bytes
cb55785 8a96218 cb55785 8a96218 2e0ebe7 8a96218 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 | ---
license: apache-2.0
pipeline_tag: text-generation
---
# D-CORE: Incentivizing Task Decomposition in Large Reasoning Models for Complex Tool Use
This repository contains the weights for **D-CORE** (**D**ecomposing tasks and **Co**mposing **Re**asoning processes), a two-stage training framework designed to enhance the task decomposition and reflective reasoning capabilities of Large Reasoning Models (LRMs) for complex tool use.
## Introduction
Effective tool use and reasoning are essential capabilities for large reasoning models (LRMs) to address complex real-world problems. Through empirical analysis, the authors identified that current LRMs lack the capability of sub-task decomposition in complex tool use scenarios, leading to "Lazy Reasoning."
To address this, D-CORE proposes a two-stage training framework:
1. **Self-distillation**: Incentivizes the LRM's task decomposition reasoning capability.
2. **Diversity-aware Reinforcement Learning (RL)**: Restores the LRM's reflective reasoning capability.
D-CORE achieves robust tool-use improvements across diverse benchmarks and model scales. Notably, D-CORE-14B establishes a new state-of-the-art on BFCLv3, outperforming 70B models despite being 5$\times$ smaller.
## Resources
- **Paper**: [D-CORE: Incentivizing Task Decomposition in Large Reasoning Models for Complex Tool Use](https://huggingface.co/papers/2602.02160)
- **Arxiv**: [2602.02160](https://arxiv.org/abs/2602.02160)
- **Code**: [EfficientAI (GitHub)](https://github.com/alibaba/EfficientAI)
## Authors
Bowen Xu, Shaoyu Wu, Hao Jiang, Kai Liu, Xin Chen, Lulu Hu, Bin Yang
## Performance
### BFCL
In our network environment, for the Web Search No Snippet task, we are unable to access certain websites (e.g., Wikipedia), which results in some deviation in the No Snippet scores.
<table>
<thead>
<tr>
<th rowspan="4">Model</th>
<th rowspan="4">Overall</th>
<th colspan="7" style="text-align:center">Agentic</th>
<th colspan="5" style="text-align:center">Multi Turn</th>
<th colspan="10" style="text-align:center">Single Turn</th>
<th colspan="2" style="text-align:center">Hallucination Measurement</th>
<th colspan="2" style="text-align:center">Format Sensitivity</th>
</tr>
<tr>
<th colspan="3" style="text-align:center">Web Search</th>
<th colspan="4" style="text-align:center">Memory</th>
<th rowspan="3">Overall Acc</th>
<th rowspan="3">Base</th>
<th rowspan="3">Miss Func</th>
<th rowspan="3">Miss Param</th>
<th rowspan="3">Long Context</th>
<th colspan="5" style="text-align:center">Non-live</th>
<th colspan="5" style="text-align:center">Live</th>
<th rowspan="3">Relevance</th>
<th rowspan="3">Irrelevance</th>
<th rowspan="3">Max Delta</th>
<th rowspan="3">SD</th>
</tr>
<tr>
<th rowspan="2">Summary</th>
<th rowspan="2">Base</th>
<th rowspan="2">No Snippet</th>
<th rowspan="2">Summary</th>
<th rowspan="2">KV</th>
<th rowspan="2">Vector</th>
<th rowspan="2">Recusive Sum</th>
<th rowspan="2">Overall Acc</th>
<th rowspan="2">Simple</th>
<th rowspan="2">Multiple</th>
<th rowspan="2">Parallel</th>
<th rowspan="2">Multiple Parallel</th>
<th rowspan="2">Overall Acc</th>
<th rowspan="2">Simple</th>
<th rowspan="2">Multiple</th>
<th rowspan="2">Parallel</th>
<th rowspan="2">Multiple Parallel</th>
</tr>
<tr>
</tr>
</thead>
<tbody>
<tr>
<td style="white-space: nowrap"><strong>D-CORE-8B</strong></td>
<td style="text-align: center">53.15</td>
<td style="text-align: center">23.00</td>
<td style="text-align: center">36.00</td>
<td style="text-align: center">10.00</td>
<td style="text-align: center">19.14</td>
<td style="text-align: center">9.03</td>
<td style="text-align: center">16.77</td>
<td style="text-align: center">31.61</td>
<td style="text-align: center">64.88</td>
<td style="text-align: center">75.50</td>
<td style="text-align: center">65.00</td>
<td style="text-align: center">60.50</td>
<td style="text-align: center">58.50</td>
<td style="text-align: center">86.85</td>
<td style="text-align: center">75.92</td>
<td style="text-align: center">92.50</td>
<td style="text-align: center">92.00</td>
<td style="text-align: center">87.00</td>
<td style="text-align: center">75.80</td>
<td style="text-align: center">78.29</td>
<td style="text-align: center">75.02</td>
<td style="text-align: center">100.00</td>
<td style="text-align: center">66.67</td>
<td style="text-align: center">75.00</td>
<td style="text-align: center">89.99</td>
<td style="text-align: center">75.0</td>
<td style="text-align: center">24.67</td>
</tr>
</tbody>
</table>
### Tau-Bench & Tau2-Bench
We use Qwen3-235B-A22B-Instruct-2507 as the user model. For each task, we sample 5 times and take the average as the final result.
<table>
<thead>
<tr>
<th rowspan="3">Model</th>
<th colspan="3" style="text-align:center">Tau-Bench</th>
<th colspan="4" style="text-align:center">Tau2-Bench</th>
</tr>
<tr>
<th rowspan="2">Overall</th>
<th rowspan="2">Retail</th>
<th rowspan="2">Airline</th>
<th rowspan="2">Overall</th>
<th rowspan="2">Retail</th>
<th rowspan="2">Airline</th>
<th rowspan="2">Telecom</th>
</tr>
<tr>
</tr>
</thead>
<tbody>
<tr>
<td style="white-space: nowrap"><strong>D-CORE-8B</strong></td>
<td style="text-align: center">44.9</td>
<td style="text-align: center">53.0</td>
<td style="text-align: center">36.8</td>
<td style="text-align: center">35.8</td>
<td style="text-align: center">43.2</td>
<td style="text-align: center">37.1</td>
<td style="text-align: center">27.2</td>
</tr>
</tbody>
</table>
### ACEBench
<table>
<thead>
<tr>
<th rowspan="3">Model</th>
<th rowspan="3">Overall</th>
<th colspan="7" style="text-align:center">Atom</th>
<th colspan="3" style="text-align:center">Single Turn</th>
<th colspan="3" style="text-align:center">Multi Turn</th>
<th rowspan="3" style="text-align:center">Similar API</th>
<th rowspan="3" style="text-align:center">Preference</th>
<th rowspan="3" style="text-align:center">Summary</th>
<th colspan="4" style="text-align:center">Special</th>
<th colspan="5" style="text-align:center">Agent</th>
</tr>
<tr>
<th rowspan="2">Summary</th>
<th rowspan="2">Bool</th>
<th rowspan="2">Enum</th>
<th rowspan="2">Number</th>
<th rowspan="2">List</th>
<th rowspan="2">Object Short</th>
<th rowspan="2">Object Deep</th>
<th rowspan="2">Summary</th>
<th rowspan="2">Singal Function</th>
<th rowspan="2">Parallel Function</th>
<th rowspan="2">Summary</th>
<th rowspan="2">Switch</th>
<th rowspan="2">Adjust</th>
<th rowspan="2">Summary</th>
<th rowspan="2">Incomplete</th>
<th rowspan="2">Error</th>
<th rowspan="2">Irrelevant</th>
<th rowspan="2">Summary</th>
<th rowspan="2">Multi Turn</th>
<th rowspan="2">Multi Turn Process</th>
<th rowspan="2">Multi Step</th>
<th rowspan="2">Multi Step Process</th>
</tr>
<tr>
</tr>
</thead>
<tbody>
<tr>
<td style="white-space: nowrap"><strong>D-CORE-8B</strong></td>
<td style="text-align: center">75.2</td>
<td style="text-align: center">82.7</td>
<td style="text-align: center">90.0</td>
<td style="text-align: center">98.0</td>
<td style="text-align: center">98.0</td>
<td style="text-align: center">98.0</td>
<td style="text-align: center">36.0</td>
<td style="text-align: center">76.0</td>
<td style="text-align: center">77.5</td>
<td style="text-align: center">85.0</td>
<td style="text-align: center">70.0</td>
<td style="text-align: center">62.0</td>
<td style="text-align: center">64.0</td>
<td style="text-align: center">60.0</td>
<td style="text-align: center">78.0</td>
<td style="text-align: center">82.0</td>
<td style="text-align: center">77.9</td>
<td style="text-align: center">78.7</td>
<td style="text-align: center">58.0</td>
<td style="text-align: center">82.0</td>
<td style="text-align: center">96.0</td>
<td style="text-align: center">59.2</td>
<td style="text-align: center">43.3</td>
<td style="text-align: center">66.8</td>
<td style="text-align: center">75.0</td>
<td style="text-align: center">80.8</td>
</tr>
</tbody>
</table>
## Citation
If you find our work useful, please cite:
```bibtex
@article{xu2026dcore,
title={D-CORE: Incentivizing Task Decomposition in Large Reasoning Models for Complex Tool Use},
author={Xu, Bowen and Wu, Shaoyu and Jiang, Hao and Liu, Kai and Chen, Xin and Hu, Lulu and Yang, Bin},
journal={arXiv preprint arXiv:2602.02160},
year={2026}
}
``` |