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}
}
```