This is the official repository for our paper "Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization". We challenge the prevailing assumption that more thinking leads to better performance in agentic search. Instead, we find that search-heavy, think-light strategies achieve superior efficiency and generalization on long-horizon web tasks. By shifting the compute budget from internal chain-of-thought reasoning toward more active web exploration, SMTL delivers competitive performance across challenging benchmarks while significantly reducing inference cost.

We fully open-source our data synthesis pipeline, data construction workflow, model weights, and inference code to support reproducible research.

---

Overview

Recent advances in deep research agents suggest that scaling reasoning depth and tool calls can substantially improve task performance. However, this reliance on linear, sequential reasoning leads to high inference costs and latency in search-intensive scenarios. Furthermore, existing agents often struggle to generalize across diverse task objectives—from deterministic question-answering with clear ground-truth answers to open-ended research problems requiring comprehensive synthesis.

To balance long-horizon search performance and computational efficiency, we propose:

• Search More, Think Less (SMTL): A unified agentic framework that replaces sequential reasoning with parallel task decomposition and concurrent tool execution, utilizing plan-driven context management for efficient long-horizon inference under constrained budgets.

• Unified Data Synthesis Pipeline: An automated data pipeline that constructs representative multi-type search tasks spanning both deterministic and open-ended settings, reducing redundant samples and teaching generalized research behavior.

SOTA Performance

Trained end-to-end using supervised fine-tuning (SFT) and reinforcement learning (RL), our SMTL agent achieves SOTA performance across multiple deep search and deep research benchmarks, while drastically improving computational efficiency.

Superior Accuracy: SMTL achieves strong, state-of-the-art results across diverse evaluation settings, including 48.6% on BrowseComp, 75.7% on GAIA, 82.0% on Xbench, and 45.9% on DeepResearch Bench.

Exceptional Efficiency: Compared to baselines like Mirothinker-v1.0, SMTL substantially reduces the average number of reasoning steps by up to 78% on BrowseComp and cuts inference latency by up to 2.6×, all while improving overall accuracy.

We fully open-source our data synthesis pipeline, data construction workflow, model weights, and inference code to ensure the reproducibility of our results. For more details, please refer to our Technical Report.

---

Quick Feature Summary

Feature Category	Supported Capabilities
Inference Pipeline	✅ Long-horizon web-agent inference ✅ Multi-step reasoning with tool calls ✅ Background launch via start_infer.sh
Tool Integration	✅ Web search and crawling servers ✅ Service status / restart / log utilities
Deployment	✅ vLLM model deployment script ✅ Single-server default setup
Configuration	✅ CLI overrides for core inference hyperparameters ✅ Environment-variable-based endpoint and key config
Statistics	✅ Per-result .output_stats.txt generation

---

Table of Contents

• Overview
• SOTA Performance
• Quick Feature Summary
• Table of Contents
• Project Layout
• Running Examples
  • 1. Environment Preparation
  • 2. Download Model Checkpoint
  • 3. Tool Server Deployment
  • 4. Model Service Deployment (vLLM)
  • 5. Inference Launch (Recommended)
  • 6. Runtime Parameters in start_infer.sh
  • 7. Single-Question Quick Test
• Configuration Notes
• FAQ
• Acknowledgement
• Related Work
• Citation
• Star

---

Project Layout

SMTL-main/
├─ .env                          ← fill in your API keys here
├─ .env.example                  ← template (copy to .env)
├─ data_synthesis/
├─ data_workflow/
├─ deploy/
│  ├─ tool_servers/
│  │  ├─ start_servers.sh
│  │  └─ scripts/
│  └─ vllm_model/
│     └─ model_deploy.sh
├─ inference/
│  ├─ start_infer.sh
│  ├─ inference_web_agent.py
│  ├─ cal_stats.py
│  ├─ benchmarks/
│  └─ logs/
├─ model/
│  └─ smtl
└─ requirements.txt

---

Running Examples

1. Environment Preparation

Create and activate a conda environment:

conda create -n smtl python=3.11 -y
conda activate smtl

Install all dependencies:

pip install -r requirements.txt

---

2. Download Model Checkpoint

Download the SMTL-30B model from Hugging Face into the model/smtl directory:

mkdir -p model/smtl
huggingface-cli download PersonalAILab/SMTL-30B --local-dir model/smtl

Or with git lfs:

git lfs install
git clone https://huggingface.co/PersonalAILab/SMTL-30B model/smtl

Model page: https://huggingface.co/PersonalAILab/SMTL-30B

2.1 Download Tokenizer

To ensure the agent functions correctly, you must download the Qwen3-4B model (used for the tokenizer) into the designated local directory:

mkdir -p /home/notebook/code/group/cjy/AFM/tokenizer/qwen3/Qwen3-4B

huggingface-cli download Qwen/Qwen3-4B \
  --local-dir /home/notebook/code/group/cjy/AFM/tokenizer/qwen3/Qwen3-4B \
  --local-dir-use-symlinks False

---

3. Tool Server Deployment

start_servers.sh loads configuration from the project root .env.
At minimum, set SERVER_HOST, CRAWL_PAGE_PORT, and WEBSEARCH_PORT in that file.

Start services:

bash deploy/tool_servers/start_servers.sh start

Useful commands:

bash deploy/tool_servers/start_servers.sh status
bash deploy/tool_servers/start_servers.sh stop
bash deploy/tool_servers/start_servers.sh restart
bash deploy/tool_servers/start_servers.sh log all

---

4. Model Service Deployment (vLLM)

bash deploy/vllm_model/model_deploy.sh

Default serving target:

• model: smtl
• host: 0.0.0.0
• port: 1

---

5. Inference Launch (Recommended)

After tool servers and model service are ready, launch inference:

bash inference/start_infer.sh

This command will:

• load .env from project root
• run inference in the background (nohup)
• write logs to inference/logs/infer_log.log by default

---

6. Runtime Parameters in start_infer.sh

Override runtime config without modifying Python files:

# Single URL
bash inference/start_infer.sh \
  --model-url http://0.0.0.0:1/v1 \
  --max-steps 100 \
  --benchmark browsecomp \
  --parallel 4 \
  --web-topk 20 \
  --log-file inference/logs/infer_log.log

# Multiple URLs (round-robin load balancing)
bash inference/start_infer.sh \
  --model-url "http://0.0.0.0:1/v1,http://0.0.0.0:2/v1" \
  --parallel 8

Supported options:

Option	Description	Default
--model-url <url(s)>	Model base URL(s); comma-separated for round-robin across multiple instances	http://0.0.0.0:1/v1
--max-steps <int>	Max reasoning steps per query	100
--benchmark <name>	Benchmark name (see available list below)	browsecomp
--parallel <int>	Parallel workers	4
--web-topk <int>	web_search top-k results	20
--log-file <path>	Log file path	inference/logs/infer_log.log

Available benchmarks: browsecomp, gaia, xbench, webwalker, frames, seal_0

For each inference result file xxx.jsonl, judge result statistics are automatically generated as:

• xxx.output_stats.txt

---

Single-Question Quick Test

To verify your deployment with a single question (without running a full benchmark), use the quick test script:

Interactive mode (prompts for input)

python inference/test.py

Pass question directly via CLI

python inference/test.py --question "Who won the 2024 Nobel Prize in Physics?"

Override model URL at runtime

python inference/test.py
--model-url http://0.0.0.0:1/v1
--question "What is the capital of France?"

The script prints the model's answer, total reasoning steps, and elapsed time. It suppresses verbose inference logs by default — set logging.WARNING → logging.INFO inside the script to enable debug output.

Prerequisites: tool servers and model service must already be running, and .env must be filled in.

---

Configuration Notes

.env (project root)

Copy .env.example to .env and fill in your credentials:

cp .env.example .env

Then edit .env and replace all placeholder values with your own keys.

Required API keys to run this project:

Key	Where to obtain
JINA_API_KEY	https://jina.ai — used by the CrawlPage tool server to fetch web page content
WEB_SEARCH_SERPER_API_KEY	https://serper.dev — used by the WebSearch tool server for web search
API_KEY	Your judge model API key (e.g. DeepSeek) — used to evaluate agent responses
BASE_URL	API endpoint for the judge model

⚠️ The project cannot run without valid JINA_API_KEY and WEB_SEARCH_SERPER_API_KEY. Please obtain these keys before starting the tool servers.

Inference Script Entry

Use inference/start_infer.sh as the standard entrypoint for all inference experiments.

---

FAQ

Inference runs but output is invalid or empty. What should I check?

1. Tool servers are up: bash deploy/tool_servers/start_servers.sh status
2. vLLM model service is reachable at the configured URL
3. URLs in .env match actual host/ports

Where are logs?

• Inference: inference/logs/infer_log.log
• Tool servers: deploy/tool_servers/logs/
• Model serving: deploy/vllm_model/logs/

---

Related Work

Listed below are friendly links to relevant agents works from OPPO PersonalAI Lab:

• Adaptive_Agent_Foundation_Models: An Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning
• Flash-Searcher: Fast and Effective Web Agents via DAG-Based Parallel Execution
• Agent Foundation Models: Chain-of-Agents: End-to-End Agent Foundation Models via Multi-Agent Distillation and Agentic RL
• TaskCraft: Automated Generation of Agentic Tasks
• OAgents: An Empirical Study of Building Effective Agents
• Agent-KB: Leveraging Cross-Domain Experience for Agentic Problem Solving
• MiCoTA: Bridging the Learnability Gap with Intermediate CoT and Teacher Assistants

Citation

If you find SMTL useful in your research or applications, please consider citing our work:

@misc{chen2026searchmorethinkless,
      title={Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization},
      author={Qianben Chen and Tianrui Qin and King Zhu and Qiexiang Wang and Chengjun Yu and Shu Xu and Jiaqi Wu and Jiayu Zhang and Xinpeng Liu and Xin Gui and Jingyi Cao and Piaohong Wang and Dingfeng Shi and He Zhu and Tiannan Wang and Yuqing Wang and Maojia Song and Tianyu Zheng and Ge Zhang and Jian Yang and Jiaheng Liu and Minghao Liu and Yuchen Eleanor Jiang and Wangchunshu Zhou},
      year={2026},
      eprint={2602.22675},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.22675},
}

---

Star