470 lines
16 KiB
Markdown
470 lines
16 KiB
Markdown
---
|
||
license: apache-2.0
|
||
language:
|
||
- en
|
||
- zh
|
||
base_model: Qwen/Qwen3-8B
|
||
library_name: transformers
|
||
tags:
|
||
- deep-research
|
||
- agent
|
||
- information-seeking
|
||
- web-search
|
||
- verification
|
||
- react
|
||
pipeline_tag: text-generation
|
||
model-index:
|
||
- name: Marco-DeepResearch-8B
|
||
results:
|
||
- task:
|
||
type: question-answering
|
||
name: BrowseComp
|
||
dataset:
|
||
name: BrowseComp
|
||
type: browsecomp
|
||
metrics:
|
||
- type: accuracy
|
||
value: 31.4
|
||
name: Accuracy
|
||
- task:
|
||
type: question-answering
|
||
name: BrowseComp-ZH
|
||
dataset:
|
||
name: BrowseComp-ZH
|
||
type: browsecomp-zh
|
||
metrics:
|
||
- type: accuracy
|
||
value: 47.1
|
||
name: Accuracy
|
||
- task:
|
||
type: question-answering
|
||
name: GAIA (text-only)
|
||
dataset:
|
||
name: GAIA
|
||
type: gaia
|
||
metrics:
|
||
- type: accuracy
|
||
value: 69.9
|
||
name: Accuracy
|
||
- task:
|
||
type: question-answering
|
||
name: xBench-DeepSearch-2505
|
||
dataset:
|
||
name: xBench-DeepSearch-2505
|
||
type: xbench-deepsearch
|
||
metrics:
|
||
- type: accuracy
|
||
value: 82.0
|
||
name: Accuracy
|
||
- task:
|
||
type: question-answering
|
||
name: WebWalkerQA
|
||
dataset:
|
||
name: WebWalkerQA
|
||
type: webwalkerqa
|
||
metrics:
|
||
- type: accuracy
|
||
value: 69.6
|
||
name: Accuracy
|
||
---
|
||
|
||
# Marco-DeepResearch-8B
|
||
|
||
<p align="center">
|
||
<a href="https://arxiv.org/abs/2603.28376">📄 Paper</a> •
|
||
<a href="https://github.com/AIDC-AI/Marco-DeepResearch/tree/main/Marco-DeepResearch-Family/Marco-Agent-DeepResearch">💻 GitHub</a>
|
||
</p>
|
||
|
||
## Introduction
|
||
|
||
**Marco DeepResearch** is an efficient 8B-scale deep research agent developed by **Alibaba International Digital Commerce (AIDC-AI)**. It autonomously conducts open-ended investigations by integrating complex information retrieval with multi-step reasoning across diverse web sources.
|
||
|
||
Marco DeepResearch is optimized through a **verification-centric framework** at three levels:
|
||
|
||
1. **Verified Data Synthesis** — Graph-based and agent-based QA synthesis with explicit verification to control difficulty and ensure answer uniqueness/correctness.
|
||
2. **Verification-Driven Trajectory Construction** — Multi-agent framework (main agent + search sub-agent + verifier sub-agent) that injects explicit verification patterns into training trajectories.
|
||
3. **Verifier-Guided Test-Time Scaling** — Uses the agent itself as a verifier at inference time, achieving +12.1 avg. improvement on benchmarks.
|
||
|
||
Under a maximum budget of 600 tool calls, Marco DeepResearch significantly outperforms 8B-scale agents and surpasses or approaches several 30B-scale agents (e.g., Tongyi DeepResearch-30B) on challenging benchmarks.
|
||
|
||
## Model Details
|
||
|
||
| Attribute | Details |
|
||
|---|---|
|
||
| **Base Model** | [Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) |
|
||
| **Parameters** | ~8B |
|
||
| **Context Window** | 128K tokens (extended via YaRN) |
|
||
| **Training** | SFT + RL (GRPO) |
|
||
| **Training Hardware** | 64 × NVIDIA A100 GPUs |
|
||
| **Max Generation Length** | 16,384 tokens |
|
||
| **Decoding** | Temperature 0.7, Top-p 0.95 |
|
||
| **Max Tool Calls** | 600 (evaluation budget) |
|
||
|
||
## Prompt Format
|
||
|
||
### Prompt Structure
|
||
|
||
The prompt follows a **System + User** two-part design (similar to OpenAI native function calling):
|
||
|
||
- **System Prompt** = Role definition + Output format + Current date + Tool definitions
|
||
- **User Prompt** = Only the user question
|
||
|
||
This ensures tool definitions stay at the front of context and won't be diluted by long multi-turn conversations.
|
||
|
||
### System Prompt Template
|
||
|
||
Below is the complete system prompt. Replace `{current_date}` with the actual date and `{tools_json}` with your tool definitions.
|
||
|
||
```
|
||
You are an expert web researcher. Your task is to find accurate, complete answers through iterative search, extraction, and verification.
|
||
|
||
## Core Principles
|
||
|
||
1) Strategic Planning
|
||
- Decompose complex questions into targeted sub-tasks
|
||
- Choose the right tool for each step
|
||
- Refine your approach based on what you learn
|
||
|
||
2) Precise Execution
|
||
- Define clear objectives before using any tool
|
||
- Provide sufficient detail for accurate results
|
||
- Avoid vague or overly broad requests
|
||
|
||
3) Rigorous Verification
|
||
- Cross-check important facts across multiple sources
|
||
- Resolve conflicts by gathering additional evidence
|
||
- Only conclude when evidence is sufficient and consistent
|
||
|
||
## Output Format
|
||
|
||
In each turn, you can either call a tool or provide the final answer.
|
||
|
||
**Call a tool:**
|
||
<think>your reasoning process</think>
|
||
<tool_call>
|
||
{"name": "tool_name", "arguments": {"param1": "value1", "param2": "value2"}}
|
||
</tool_call>
|
||
|
||
**Provide final answer (when you have gathered enough information):**
|
||
<think>your reasoning and analysis</think>
|
||
<answer>the direct answer to the question</answer>
|
||
|
||
Note: All reasoning should be in <think>, <answer> should contain only the final answer.
|
||
|
||
Current date: {current_date}
|
||
|
||
# Tools
|
||
|
||
You may call one or more functions to assist with the user query.
|
||
|
||
You are provided with function signatures within <tools></tools> XML tags:
|
||
<tools>
|
||
{tools_json}
|
||
</tools>
|
||
|
||
For each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:
|
||
<tool_call>
|
||
{"name": <function-name>, "arguments": <args-json-object>}
|
||
</tool_call>
|
||
```
|
||
|
||
### Tool Definition Format
|
||
|
||
Tools use the **OpenAI function calling format** and are placed inside the system prompt. Example:
|
||
|
||
```json
|
||
[
|
||
{
|
||
"type": "function",
|
||
"function": {
|
||
"name": "search",
|
||
"description": "Search the web via Google to find relevant information and URLs.",
|
||
"parameters": {
|
||
"type": "object",
|
||
"properties": {
|
||
"querys": {
|
||
"type": "array",
|
||
"items": {"type": "string"},
|
||
"description": "Search queries for finding relevant information. Supports single or multiple queries."
|
||
}
|
||
},
|
||
"required": ["querys"]
|
||
}
|
||
}
|
||
},
|
||
{
|
||
"type": "function",
|
||
"function": {
|
||
"name": "visit",
|
||
"description": "Read webpage content to extract specific information, verify claims, or understand context.",
|
||
"parameters": {
|
||
"type": "object",
|
||
"properties": {
|
||
"urls": {
|
||
"type": "array",
|
||
"items": {"type": "string"},
|
||
"description": "URL(s) to visit. Supports single or multiple urls."
|
||
},
|
||
"goal": {
|
||
"type": "string",
|
||
"description": "The specific information to retrieve. Be precise, not vague."
|
||
}
|
||
},
|
||
"required": ["urls", "goal"]
|
||
}
|
||
}
|
||
}
|
||
]
|
||
```
|
||
|
||
### Model Output Format
|
||
|
||
In each turn, the model produces one of two structured outputs:
|
||
|
||
**Tool call turn:**
|
||
```xml
|
||
<think>
|
||
I need to search for information about X to answer the user's question.
|
||
Let me start by searching for...
|
||
</think>
|
||
<tool_call>
|
||
{"name": "search", "arguments": {"querys": ["search query here"]}}
|
||
</tool_call>
|
||
```
|
||
|
||
**Final answer turn:**
|
||
```xml
|
||
<think>
|
||
Based on the evidence gathered from multiple sources, I can now conclude that...
|
||
Let me verify: Source A says X, Source B confirms X, and Source C also supports X.
|
||
</think>
|
||
<answer>
|
||
The direct answer to the question.
|
||
</answer>
|
||
```
|
||
|
||
### Multi-Turn Conversation Format
|
||
|
||
A complete multi-turn agent session looks like:
|
||
|
||
```
|
||
[System] Role + Tools + Date
|
||
|
||
[User] What is the population of the largest city in Switzerland?
|
||
|
||
[Asst] <think>I need to find the largest city in Switzerland first...</think>
|
||
<tool_call>{"name": "search", "arguments": {"querys": ["largest city in Switzerland"]}}</tool_call>
|
||
|
||
[User] <tool_response>{"results": [{"title": "Zürich - Wikipedia", ...}]}</tool_response>
|
||
|
||
[Asst] <think>Zürich is the largest city. Now let me find its population...</think>
|
||
<tool_call>{"name": "visit", "arguments": {"urls": ["..."], "goal": "population of Zürich"}}</tool_call>
|
||
|
||
[User] <tool_response>{"found": true, "content": "Population: 434,335 (2024)"}</tool_response>
|
||
|
||
[Asst] <think>I found the answer with a reliable source...</think>
|
||
<answer>The largest city in Switzerland is Zürich, with a population of approximately 434,335.</answer>
|
||
```
|
||
|
||
## Quick Start
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
import json
|
||
from datetime import datetime
|
||
|
||
model_name = "AIDC-AI/Marco-DeepResearch-8B"
|
||
|
||
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
model_name,
|
||
torch_dtype="auto",
|
||
device_map="auto",
|
||
trust_remote_code=True,
|
||
)
|
||
|
||
# 1. Define tools (OpenAI function calling format)
|
||
tools = [
|
||
{
|
||
"type": "function",
|
||
"function": {
|
||
"name": "search",
|
||
"description": "Search the web via Google to find relevant information and URLs.",
|
||
"parameters": {
|
||
"type": "object",
|
||
"properties": {
|
||
"querys": {
|
||
"type": "array",
|
||
"items": {"type": "string"},
|
||
"description": "Search queries for finding relevant information. Supports single or multiple queries."
|
||
}
|
||
},
|
||
"required": ["querys"]
|
||
}
|
||
}
|
||
},
|
||
{
|
||
"type": "function",
|
||
"function": {
|
||
"name": "visit",
|
||
"description": "Read webpage content to extract specific information, verify claims, or understand context.",
|
||
"parameters": {
|
||
"type": "object",
|
||
"properties": {
|
||
"urls": {
|
||
"type": "array",
|
||
"items": {"type": "string"},
|
||
"description": "URL(s) to visit. Supports single or multiple urls."
|
||
},
|
||
"goal": {
|
||
"type": "string",
|
||
"description": "The specific information to retrieve. Be precise, not vague."
|
||
}
|
||
},
|
||
"required": ["urls", "goal"]
|
||
}
|
||
}
|
||
}
|
||
]
|
||
|
||
# 2. Build system prompt
|
||
ROLE_PROMPT = """You are an expert web researcher. Your task is to find accurate, complete answers through iterative search, extraction, and verification.
|
||
|
||
## Core Principles
|
||
|
||
1) Strategic Planning
|
||
- Decompose complex questions into targeted sub-tasks
|
||
- Choose the right tool for each step
|
||
- Refine your approach based on what you learn
|
||
|
||
2) Precise Execution
|
||
- Define clear objectives before using any tool
|
||
- Provide sufficient detail for accurate results
|
||
- Avoid vague or overly broad requests
|
||
|
||
3) Rigorous Verification
|
||
- Cross-check important facts across multiple sources
|
||
- Resolve conflicts by gathering additional evidence
|
||
- Only conclude when evidence is sufficient and consistent
|
||
|
||
## Output Format
|
||
|
||
In each turn, you can either call a tool or provide the final answer.
|
||
|
||
**Call a tool:**
|
||
<think>your reasoning process</think>
|
||
<tool_call>
|
||
{"name": "tool_name", "arguments": {"param1": "value1", "param2": "value2"}}
|
||
</tool_call>
|
||
|
||
**Provide final answer (when you have gathered enough information):**
|
||
<think>your reasoning and analysis</think>
|
||
<answer>the direct answer to the question</answer>
|
||
|
||
Note: All reasoning should be in <think>, <answer> should contain only the final answer."""
|
||
|
||
current_date = datetime.now().strftime("%Y-%m-%d")
|
||
tools_json = "\n".join([json.dumps(t, ensure_ascii=False) for t in tools])
|
||
|
||
system_prompt = f"""{ROLE_PROMPT}
|
||
|
||
Current date: {current_date}
|
||
|
||
# Tools
|
||
|
||
You may call one or more functions to assist with the user query.
|
||
|
||
You are provided with function signatures within <tools></tools> XML tags:
|
||
<tools>
|
||
{tools_json}
|
||
</tools>
|
||
|
||
For each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:
|
||
<tool_call>
|
||
{{"name": <function-name>, "arguments": <args-json-object>}}
|
||
</tool_call>"""
|
||
|
||
# 3. Build messages and generate
|
||
messages = [
|
||
{"role": "system", "content": system_prompt},
|
||
{"role": "user", "content": "Who won the 2026 Turing Award?"},
|
||
]
|
||
|
||
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
||
inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
||
outputs = model.generate(**inputs, max_new_tokens=16384, temperature=0.7, top_p=0.95)
|
||
response = tokenizer.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True)
|
||
print(response)
|
||
```
|
||
|
||
> For the full **multi-turn agent loop**, see the [GitHub repository](https://github.com/AIDC-AI/Marco-DeepResearch/tree/main/Marco-DeepResearch-Family/Marco-Agent-DeepResearch).
|
||
|
||
## Benchmark Results
|
||
|
||
Evaluated on a suite of deep search benchmarks under a maximum budget of **600 tool calls**.
|
||
|
||
<p align="center">
|
||
<img src="https://raw.githubusercontent.com/AIDC-AI/Marco-DeepResearch/refs/heads/main/Marco-DeepResearch-Family/Marco-Agent-DeepResearch/assets/benchmark_chart_v2.png" alt="Marco DeepResearch benchmark performance across BrowseComp, BrowseComp-ZH, xBench-DeepSearch-2510, and GAIA (text-only)" width="100%" />
|
||
</p>
|
||
|
||
| Model | BrowseComp | BrowseComp-ZH | GAIA (text-only) | WebWalkerQA | xBench-DS-2505 | xBench-DS-2510 | DeepSearchQA | HLE-Text |
|
||
|---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
|
||
| *Foundation Models with Tools* | | | | | | | | |
|
||
| GLM-4.7 | 67.5 | 66.6 | 61.9 | – | 72.0 | 52.3 | – | 42.8 |
|
||
| Minimax-M2.1 | 62.0 | 47.8 | 64.3 | – | 68.7 | 43.0 | – | 19.1 |
|
||
| DeepSeek-V3.2 | 67.6 | 65.0 | 75.1 | – | 78.0 | 55.7 | 60.9 | 40.8 |
|
||
| Kimi-K2.5 | 74.9 | 62.3 | – | – | – | 46.0 | 77.1 | – |
|
||
| Claude-4-Sonnet | 12.2 | 29.1 | 68.3 | 61.7 | 64.6 | – | – | – |
|
||
| Claude-4.5-Opus | 67.8 | 62.4 | – | – | – | – | 80.0 | – |
|
||
| OpenAI-o3 | 49.7 | 58.1 | – | 71.7 | 67.0 | – | – | – |
|
||
| OpenAI GPT-5 High | 54.9 | 65.0 | 76.4 | – | 77.8 | 75.0 | 79.0 | – |
|
||
| Gemini-3.0-Pro | 59.2 | 66.8 | – | – | – | 53.0 | 76.9 | – |
|
||
| *Trained Agents (≥30B)* | | | | | | | | |
|
||
| MiroThinker-v1.7-mini | 67.9 | 72.3 | 80.3 | – | – | 57.2 | 67.9 | 36.4 |
|
||
| MiroThinker-v1.5-235B | 69.8 | 71.5 | 80.8 | – | 77.1 | – | – | 39.2 |
|
||
| MiroThinker-v1.5-30B | 56.1 | 66.8 | 72.0 | – | 73.1 | – | – | 31.0 |
|
||
| MiroThinker-v1.0-72B | 47.1 | 55.6 | 81.9 | 62.1 | 77.8 | – | – | 37.7 |
|
||
| MiroThinker-v1.0-30B | 41.2 | 47.8 | 73.5 | 61.0 | 70.6 | – | – | 33.4 |
|
||
| SMTL-30B-300 | 48.6 | – | 75.7 | 76.5 | 82.0 | – | – | – |
|
||
| Tongyi-DR-30B | 43.4 | 46.7 | 70.9 | 72.2 | 75.0 | 55.0 | – | 32.9 |
|
||
| WebSailor-V2-30B | 35.3 | 44.1 | 74.1 | – | 73.7 | – | – | – |
|
||
| DeepMiner-32B-RL | 33.5 | 40.1 | 58.7 | – | 62.0 | – | – | – |
|
||
| OpenSeeker-30B-SFT | 29.5 | 48.4 | – | – | 74.0 | – | – | – |
|
||
| *Trained Agents (≤8B)* | | | | | | | | |
|
||
| AgentCPM-Explore-4B | 24.1 | 29.1 | 63.9 | 68.1 | 70.0 | 34.0* | 32.8* | 19.1 |
|
||
| WebExplorer-8B-RL | 15.7 | 32.0 | 50.0 | 62.7 | 53.7 | 23.0* | 17.8* | 17.3 |
|
||
| RE-TRAC-4B | 30.0 | 36.1 | 70.4 | – | 76.6 | – | – | 22.2 |
|
||
| MiroThinker-v1.0-8B | 31.1 | 40.2 | 66.4 | 60.6 | 60.6 | 34.0* | 36.7* | 21.5 |
|
||
| **Marco-DR-8B (Ours)** | **31.4** | **47.1** | **69.9** | **69.6** | **82.0** | **42.0** | **29.9** | **22.5** |
|
||
|
||
> \* marks scores we reproduced with our own implementation; other scores are from the respective official reports.
|
||
|
||
## Intended Use
|
||
|
||
Marco DeepResearch is designed for:
|
||
- **Open-ended web research** — Autonomously investigating complex questions across multiple web sources
|
||
- **Multi-hop question answering** — Solving questions that require chaining information from multiple documents
|
||
- **Information seeking** — Navigating the web to locate specific, hard-to-find information
|
||
- **Fact verification** — Verifying claims through multi-source evidence aggregation
|
||
|
||
## Limitations
|
||
|
||
- Performance depends on the quality and accessibility of external web search tools and APIs.
|
||
- The model is optimized for information-seeking tasks and may not generalize to all types of reasoning tasks.
|
||
- Enabling test-time scaling introduces extra inference overhead; the number of tool-call rounds can be tuned based on the use case.
|
||
|
||
## Citation
|
||
|
||
```bibtex
|
||
@article{zhu2026marco,
|
||
title={Marco DeepResearch: Unlocking Efficient Deep Research Agents via Verification-Centric Design},
|
||
author={Bin Zhu and Qianghuai Jia and Tian Lan and Junyang Ren and Feng Gu and Feihu Jiang and Longyue Wang and Zhao Xu and Weihua Luo},
|
||
journal={arXiv preprint arXiv:2603.28376},
|
||
year={2026}
|
||
}
|
||
```
|
||
|
||
## License
|
||
|
||
This model is released under the [Apache 2.0 License](LICENSE).
|