Hugging Face's logo

BAAI
/
AREX-2

Image-Text-to-Text
Transformers
Safetensors
qwen3_5
agent
deep-research
reasoning
tool-use
long-context
self-improvement
conversational
Model card Files
xet
 Community
5
AREX-2
File size: 14,728 Bytes
72c9de8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d4e3502
72c9de8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d4e3502
 
 
 
 
 
72c9de8
 
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
---
license: apache-2.0
library_name: transformers
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3.8-27B
tags:
  - qwen3_5
  - agent
  - deep-research
  - reasoning
  - tool-use
  - long-context
  - self-improvement
  - safetensors
---

<div align="center">
  <img src="assets/arex-logo.png" width="40%" alt="AREX" style="display:block;margin:0 auto -2px;">
  <strong>AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks</strong><br>
  <a href="https://arxiv.org/abs/2609.38288"><img src="https://img.shields.io/badge/-Paper-B31B1B?style=for-the-badge&amp;logo=arxiv&amp;logoColor=white" alt="Paper"></a>
  <a href="https://github.com/VectorSpaceLab/AREX-2"><img src="https://img.shields.io/badge/-Homepage-24292F?style=for-the-badge&amp;logo=github&amp;logoColor=white" alt="Homepage"></a>
</div>

## Introduction

AREX-2 is a 27B-parameter long-horizon agent model from the Beijing Academy of
Artificial Intelligence (BAAI). It learns to improve a solution over multiple
test-time rounds: propose, measure, reflect, and revise.

AREX-2 is trained on machine-learning and algorithmic-programming tasks with
verifiable feedback, together with the existing AREX deep-research data. The
learned self-improvement behavior transfers to deep research without adding new
search trajectories.

- **Architecture:** Dense Qwen3.8-compatible multimodal model
- **Parameters:** 27B
- **Context length:** 262,144 tokens

## Key features

- **Long-horizon self-improvement:** turns extra test-time rounds into useful
  solution refinement.
- **Feedback-driven reflection:** reads scores, logs, errors, and timings to
  decide what to change next.
- **Cross-domain performance:** training on coding and machine-learning tasks
  also improves the model's deep-research performance.
- **Long-horizon reasoning:** sustains productive iteration as the task budget
  grows.

## Evaluation

AREX-2 is evaluated on algorithmic programming, machine-learning engineering,
deep research, and general agentic reasoning. Results follow the protocols
reported in the AREX-2 paper.

### Coding and machine-learning engineering

<table cellpadding="7" style="display:table!important;overflow:visible!important;width:auto;max-width:100%;border:0!important;border-radius:0!important;border-collapse:collapse!important;font-size:13px;table-layout:auto">
<thead>
<tr style="color:#0F766E;border-bottom:2px solid #0F766E">
<th style="padding:8px;text-align:left">Model</th>
<th style="padding:8px;text-align:center">Params</th>
<th style="padding:8px;text-align:center">Frontier-CS</th>
<th style="padding:8px;text-align:center">MLE-Lite</th>
</tr>
</thead>
<tbody>
<tr><td colspan="4" style="padding:8px 10px;font-weight:600;color:#0F766E;background:rgba(15,118,110,.1)">Closed-weight models</td></tr>
<tr><td>GPT-5.6 Sol</td><td style="text-align:center">-</td><td style="text-align:center">76.4</td><td style="text-align:center">72.7</td></tr>
<tr><td>Claude Opus 4.8</td><td style="text-align:center">-</td><td style="text-align:center">74.5</td><td style="text-align:center">63.6</td></tr>
<tr><td>GPT-5.5</td><td style="text-align:center">-</td><td style="text-align:center">72.1</td><td style="text-align:center">68.2</td></tr>
<tr><td>Gemini-3.1-Pro</td><td style="text-align:center">-</td><td style="text-align:center">68.9</td><td style="text-align:center">-</td></tr>
<tr><td>Qwen3.7-Max</td><td style="text-align:center">-</td><td style="text-align:center">61.9</td><td style="text-align:center">-</td></tr>
<tr><td colspan="4" style="padding:8px 10px;font-weight:600;color:#0F766E;background:rgba(15,118,110,.1)">Open-weight models</td></tr>
<tr><td>Kimi-K3</td><td style="text-align:center">2.8T</td><td style="text-align:center">-</td><td style="text-align:center">72.7</td></tr>
<tr><td>Naive-N0.5-Flash</td><td style="text-align:center">309B</td><td style="text-align:center">-</td><td style="text-align:center">73.7</td></tr>
<tr><td>DeepSeek-V4-Pro</td><td style="text-align:center">1.6T</td><td style="text-align:center">44.7</td><td style="text-align:center">54.5</td></tr>
<tr><td>DeepSeek-V4-Flash</td><td style="text-align:center">284B</td><td style="text-align:center">39.1</td><td style="text-align:center">51.5</td></tr>
<tr><td>Kimi-K2.7-Code</td><td style="text-align:center">1T</td><td style="text-align:center">54.7</td><td style="text-align:center">-</td></tr>
<tr><td>GLM-5.3-Flash</td><td style="text-align:center">320B</td><td style="text-align:center">50.4</td><td style="text-align:center">-</td></tr>
<tr><td>Kimi-K2.6</td><td style="text-align:center">1T</td><td style="text-align:center">46.9</td><td style="text-align:center">66.7</td></tr>
<tr><td>Frontis-MA1-35B</td><td style="text-align:center">35B</td><td style="text-align:center">-</td><td style="text-align:center">71.2</td></tr>
<tr><td>BigBang-V1</td><td style="text-align:center">35B</td><td style="text-align:center">-</td><td style="text-align:center">59.1</td></tr>
<tr><td>Qwen3.6-35B-A3B</td><td style="text-align:center">35B</td><td style="text-align:center">23.4</td><td style="text-align:center">39.4</td></tr>
<tr style="background:rgba(15,118,110,.08);font-weight:600;color:#0F766E"><td style="padding:7px 10px">AREX-2</td><td style="text-align:center">27B</td><td style="text-align:center">70.7</td><td style="text-align:center">81.8</td></tr>
</tbody>
</table>

<small>Frontier-CS is the 188-task Agent Track. MLE-Lite reports Any Medal
averaged over three seeds.</small>

### General agentic reasoning and deep research

<table cellpadding="7" style="display:table!important;overflow:visible!important;width:auto;max-width:100%;border:0!important;border-radius:0!important;border-collapse:collapse!important;font-size:13px;table-layout:auto">
<thead>
<tr style="color:#0F766E;border-bottom:2px solid #0F766E">
<th style="padding:8px;text-align:left">Model</th>
<th style="padding:8px;text-align:center">Params</th>
<th style="padding:8px;text-align:center">BrowseComp</th>
<th style="padding:8px;text-align:center">HLE</th>
<th style="padding:8px;text-align:center">GAIA</th>
<th style="padding:8px;text-align:center">DeepSearchQA</th>
</tr>
</thead>
<tbody>
<tr><td colspan="6" style="padding:8px 10px;font-weight:600;color:#0F766E;background:rgba(15,118,110,.1)">Frontier models</td></tr>
<tr><td>GPT-5.6 Sol</td><td style="text-align:center">-</td><td style="text-align:center">90.4</td><td style="text-align:center">58.0*</td><td style="text-align:center">-</td><td style="text-align:center">-</td></tr>
<tr><td>GPT-5.6 Terra</td><td style="text-align:center">-</td><td style="text-align:center">87.5</td><td style="text-align:center">-</td><td style="text-align:center">-</td><td style="text-align:center">-</td></tr>
<tr><td>GPT-5.6 Luna</td><td style="text-align:center">-</td><td style="text-align:center">83.3</td><td style="text-align:center">-</td><td style="text-align:center">-</td><td style="text-align:center">-</td></tr>
<tr><td>Kimi-K3</td><td style="text-align:center">2.8T</td><td style="text-align:center">91.2</td><td style="text-align:center">56.0*</td><td style="text-align:center">-</td><td style="text-align:center">95.0</td></tr>
<tr><td>Claude Fable 5</td><td style="text-align:center">-</td><td style="text-align:center">88.0</td><td style="text-align:center">64.5*</td><td style="text-align:center">-</td><td style="text-align:center">94.2</td></tr>
<tr><td>Claude Opus 4.8</td><td style="text-align:center">-</td><td style="text-align:center">84.3</td><td style="text-align:center">57.9*</td><td style="text-align:center">-</td><td style="text-align:center">93.1</td></tr>
<tr><td>GPT-5.5</td><td style="text-align:center">-</td><td style="text-align:center">84.4</td><td style="text-align:center">52.2*</td><td style="text-align:center">87.4</td><td style="text-align:center">-</td></tr>
<tr><td>Gemini-3.1-Pro</td><td style="text-align:center">-</td><td style="text-align:center">85.9</td><td style="text-align:center">51.4*</td><td style="text-align:center">80.6</td><td style="text-align:center">93.3</td></tr>
<tr><td colspan="6" style="padding:8px 10px;font-weight:600;color:#0F766E;background:rgba(15,118,110,.1)">Large models (&gt;40B)</td></tr>
<tr><td>GLM-5</td><td style="text-align:center">744B</td><td style="text-align:center">75.9</td><td style="text-align:center">50.4</td><td style="text-align:center">70.0</td><td style="text-align:center">-</td></tr>
<tr><td>Kimi-K2.6</td><td style="text-align:center">1T</td><td style="text-align:center">83.2</td><td style="text-align:center">54.0*</td><td style="text-align:center">80.6</td><td style="text-align:center">92.5</td></tr>
<tr><td>GLM-5.3-Flash</td><td style="text-align:center">320B</td><td style="text-align:center">-</td><td style="text-align:center">55.3*</td><td style="text-align:center">78.8</td><td style="text-align:center">-</td></tr>
<tr><td>DeepSeek-V4-Flash</td><td style="text-align:center">284B</td><td style="text-align:center">73.2</td><td style="text-align:center">45.1</td><td style="text-align:center">57.5</td><td style="text-align:center">90.6</td></tr>
<tr><td>DeepSeek-V4-Pro</td><td style="text-align:center">1.6T</td><td style="text-align:center">83.4</td><td style="text-align:center">48.2</td><td style="text-align:center">71.1</td><td style="text-align:center">88.7</td></tr>
<tr><td>MiroThinker-1.7</td><td style="text-align:center">235B</td><td style="text-align:center">74.0</td><td style="text-align:center">42.9</td><td style="text-align:center">82.7</td><td style="text-align:center">72.1</td></tr>
<tr><td>XYZ-Aquila-pro</td><td style="text-align:center">397B</td><td style="text-align:center">84.8</td><td style="text-align:center">53.3</td><td style="text-align:center">-</td><td style="text-align:center">92.5</td></tr>
<tr><td>Iris-pro</td><td style="text-align:center">397B</td><td style="text-align:center">88.6</td><td style="text-align:center">56.4</td><td style="text-align:center">-</td><td style="text-align:center">92.9</td></tr>
<tr><td>AREX-Base</td><td style="text-align:center">122B</td><td style="text-align:center">82.5</td><td style="text-align:center">52.4</td><td style="text-align:center">85.4</td><td style="text-align:center">89.9</td></tr>
<tr><td colspan="6" style="padding:8px 10px;font-weight:600;color:#0F766E;background:rgba(15,118,110,.1)">Small models (&le;40B)</td></tr>
<tr><td>Tongyi-DeepResearch-30B</td><td style="text-align:center">30B</td><td style="text-align:center">43.4</td><td style="text-align:center">32.9</td><td style="text-align:center">70.9</td><td style="text-align:center">-</td></tr>
<tr><td>Qwen3.5-35B</td><td style="text-align:center">35B</td><td style="text-align:center">61.0</td><td style="text-align:center">47.4</td><td style="text-align:center">80.0</td><td style="text-align:center">68.5</td></tr>
<tr><td>XYZ-Aquila-mini</td><td style="text-align:center">35B</td><td style="text-align:center">78.8</td><td style="text-align:center">51.1</td><td style="text-align:center">97.1</td><td style="text-align:center">89.5</td></tr>
<tr><td>BigBang-V1</td><td style="text-align:center">35B</td><td style="text-align:center">76.5</td><td style="text-align:center">50.3</td><td style="text-align:center">-</td><td style="text-align:center">-</td></tr>
<tr><td>Quest-35B</td><td style="text-align:center">35B</td><td style="text-align:center">64.6</td><td style="text-align:center">37.2</td><td style="text-align:center">80.8</td><td style="text-align:center">-</td></tr>
<tr><td>Apodex-1.0-mini</td><td style="text-align:center">35B</td><td style="text-align:center">71.5</td><td style="text-align:center">46.8</td><td style="text-align:center">-</td><td style="text-align:center">82.2</td></tr>
<tr><td>Agents-A1</td><td style="text-align:center">35B</td><td style="text-align:center">75.5</td><td style="text-align:center">47.6</td><td style="text-align:center">96.0</td><td style="text-align:center">-</td></tr>
<tr><td>MiroThinker-1.7-mini</td><td style="text-align:center">30B</td><td style="text-align:center">67.9</td><td style="text-align:center">36.4</td><td style="text-align:center">80.3</td><td style="text-align:center">67.9</td></tr>
<tr><td>Iris-mini</td><td style="text-align:center">35B</td><td style="text-align:center">82.2</td><td style="text-align:center">52.3</td><td style="text-align:center">-</td><td style="text-align:center">86.9</td></tr>
<tr><td>AREX-Turbo</td><td style="text-align:center">4B</td><td style="text-align:center">70.7</td><td style="text-align:center">40.6</td><td style="text-align:center">81.6</td><td style="text-align:center">78.5</td></tr>
<tr style="background:rgba(15,118,110,.12);font-weight:700;color:#0F766E"><td>AREX-2</td><td style="text-align:center">27B</td><td style="text-align:center">84.0</td><td style="text-align:center">52.6</td><td style="text-align:center">92.2</td><td style="text-align:center">93.8</td></tr>
</tbody>
</table>

<small>HLE values marked * are from the full HLE set; unmarked values use the
text-only subset.</small>

## Inference

Use a recent Transformers release with Qwen3.8 support.

```bash
pip install -U torch transformers accelerate
```

```python
import torch
from transformers import AutoModelForMultimodalLM, AutoProcessor

model_id = "BAAI/AREX-2"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForMultimodalLM.from_pretrained(
    model_id, dtype=torch.bfloat16, device_map="auto"
)

messages = [{
    "role": "user",
    "content": "Propose a solution and explain how you would improve it over several rounds.",
}]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    outputs = model.generate(**inputs, max_new_tokens=1024)
print(processor.decode(
    outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True
))
```

## Intended use

AREX-2 is intended for research on long-horizon agents, iterative problem
solving, machine-learning engineering, algorithmic coding, and tool-augmented
deep research.

## License

AREX-2 is released under the [Apache License 2.0](LICENSE). Follow the terms
and notices for the Qwen base model and any downstream data or tools.

## Citation

```bibtex
@article{2026arex2,
  title   = {AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks},
  author  = {Qian, Hongjin and Li, Chaofan and Luo, Kun and Wei, Wenqing and Chen, Jianlyu and Lu, Shuqi and Hu, Yuyang and Xiao, Hongwang and Wang, Hui and Li, Chaozhuo and Ye, Qiwei and Dou, Zhicheng and Lian, Defu and Liu, Zheng},
  journal = {arXiv preprint arXiv:2609.38288},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.38288}
}
```