SearchGym: Bootstrapping Real-World Search Agents via Cost-Effective and High-Fidelity Environment Simulation

| 📰 Paper | 🤗 Models | 🤗 Datasets |

SearchGym is a high-fidelity simulation environment designed to train robust search agents without the prohibitive costs and noise associated with live web training. By constructing a verifiable knowledge graph and an aligned document corpus, SearchGym provides a closed-loop environment where every reasoning task is factually grounded and strictly solvable.

📊 Data Construction Pipeline

🌟 Core Environments

SearchGym operates across three distinct search environments, each serving a specific purpose in the pipeline (Training vs. Evaluation).

Environment Type	Backend	Purpose	Code Identifier	Required Setup
1. Synthetic (SearchGym)	Meilisearch	Training (RL). High-speed, typo-tolerant, verifiable ground truth.	meilisearch-local	Meilisearch Binary + Mini-Wiki Data
2. Local (Wikipedia)	Pyserini / FAISS	Standard Eval (NQ, HotpotQA). Static 2018 Wiki snapshot.	async-search-access	Local RAG Server + Index Files
3. Live Web	Serper + Jina	Open-Ended Eval (GAIA, DeepSearch). Real-time web browsing.	async-web-search-access	API Keys (Serper, Jina)

---

🛠️ 1. Installation

Environment Setup

Create a conda environment and install dependencies. Note that SearchGym relies on AReaL for asynchronous RL training.

# 1. Create conda environment
conda create -n SearchGym python=3.12
conda activate SearchGym

# 2. Install dependencies
# Navigate to AReal directory (assumed submodule or copied)
cd AReal
bash examples/env/setup-pip-deps.sh

# 3. Validate installation
python examples/env/validate_installation.py

Download Data

Download the training data (Synthetic Corpus) and evaluation benchmarks.

git clone https://huggingface.co/datasets/hkuzxc/SearchGym-test-data
# Ensure the directory structure is: project_root/SearchGym-test-data/

---

⚙️ 2. Environment Configuration

A. Training Environment (Synthetic / Meilisearch)

Used for: Stage 1 & Stage 2 RL Training

1. Install & Start Meilisearch:

   # Install
   curl -L https://install.meilisearch.com | sh
   
   # Start Server (Background)
   # Master key matches config in SearchGym/meilisearch_client.py
   mkdir -p logs && nohup ./meilisearch --master-key="aSampleMasterKey" > logs/meilisearch.log 2>&1 &
   

2. Generate & Index Data (Optional if you have the JSON): If you need to regenerate the synthetic data:

   cd mini-wiki
   export DEEPSEEK_API_KEY="your-key"
   python scripts/run_all_steps.py --steps all
   

3. Push Data to Meilisearch:

   curl -X POST 'http://127.0.0.1:7700/indexes/wiki/documents?primaryKey=id' \
     -H 'Content-Type: application/json' \
     -H 'Authorization: Bearer aSampleMasterKey' \
     --data-binary @mini-wiki/outputs/wiki/wiki_with_urls.json
   

B. Local Evaluation Environment (Wikipedia / FAISS)

Used for: NQ, HotpotQA, TriviaQA, etc.

1. Setup Environment:

   conda create -n retriever python=3.10
   conda activate retriever
   
   # Install PyTorch with CUDA support
   conda install pytorch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 pytorch-cuda=12.1 -c pytorch -c nvidia
   
   # Install dependencies
   pip install transformers datasets pyserini
   # Install GPU-accelerated FAISS
   conda install -c pytorch -c nvidia faiss-gpu=1.8.0
   
   # Install API server dependencies
   pip install uvicorn fastapi
   

2. Download Indices: Download the E5 retriever index and corpus from ASearcher-Local-Knowledge.

3. Launch Retrieval Server: Modify scripts/launch_local_server.sh with your paths, then run:

   bash scripts/launch_local_server.sh 8000 /path/to/server_address_log/
   

   This starts a FastAPI server that acts as the search engine.

C. Web Evaluation Environment (Live)

Used for: GAIA, xBench-DeepSearch

This environment requires external API keys. No local server is needed, but configuration files must be updated.

---

📝 3. Configuration Files

Training Configs

Located in SearchGym/SearchGym/configs/. Example: SearchGym_stage1.yaml

# ... (Cluster settings)

# Model Path
actor:
  path: /path/to/base/model # e.g., Qwen2.5-3B

# Environment Selection
# "meilisearch-local" points to the setup in Section 2A
search_client_type: meilisearch-local 

# Concurrency & Queue
use_queue: true
redis_config:
  url: "redis://localhost:6379"

# Dataset Paths (Relative to project root)
train_dataset:
  path: ../SearchGym-test-data/mini_wiki_train/stage1/stage1_train.jsonl

Evaluation Config

Located in SearchGym/evaluation/eval_config.yaml.

api_keys:
  # For Web Env (GAIA/xBench)
  serper_api_key: "your-serper-api-key"
  jina_api_key: "your-jina-api-key"
  
settings:
  # For Local Env (NQ/HotpotQA)
  # Matches the IP/Port from Section 2B
  local_server:
    address: "127.0.0.1"
    port: "8000"

---

🚀 4. Training

We use a curriculum learning approach. Ensure SEARCHGYM_ROOT and WANDB_API_KEY are set in the scripts.

Stage 1: Foundational Skill Acquisition

cd SearchGym
bash run_SearchGym_stage1.sh

Stage 2: Advanced Reasoning Development Update run_SearchGym_stage2.sh to point to the checkpoint from Stage 1.

bash run_SearchGym_stage2.sh

---

📊 5. Evaluation

We provide scripts for both Local and Online evaluations.

Batch Evaluation (Recommended)

Local Benchmarks (Bamboogle, NQ, etc.): Edit SearchGym/evaluation/batch_run_eval_local.sh:

• AGENT_TYPE=SearchGym
• SEARCH_CLIENT_TYPE=async-search-access (Uses Local RAG Server)

cd SearchGym/evaluation
bash batch_run_eval_local.sh

Online Benchmarks (GAIA, xBench): Edit SearchGym/evaluation/batch_run_eval_online.sh:

• AGENT_TYPE=SearchGym
• SEARCH_CLIENT_TYPE=async-web-search-access (Uses Serper/Jina)

cd SearchGym/evaluation
bash batch_run_eval_online.sh

---

📦 Pre-trained Models

We provide pre-trained SearchGym models on HuggingFace: SearchGym Collection

Model	Size	Link
SearchGym_Qwen_2.5_3B_Base	3B	hkuzxc/SearchGym_Qwen_2.5_3B_Base
SearchGym_Qwen_2.5_3B_Instruct	3B	hkuzxc/SearchGym_Qwen_2.5_3B_Instruct
SearchGym_Qwen_2.5_7B_Base	7B	hkuzxc/SearchGym_Qwen_2.5_7B_Base
SearchGym_Qwen_2.5_7B_Instruct	7B	hkuzxc/SearchGym_Qwen_2.5_7B_Instruct
SearchGym_Qwen_3_4B	4B	hkuzxc/SearchGym_Qwen_3_4B
SearchGym_Qwen_3_8B	8B	hkuzxc/SearchGym_Qwen_3_8B
SearchGym_Llama_3.2_3B_Instruct	3B	hkuzxc/SearchGym_Llama_3.2_3B_Instruct

📚 Citation

@misc{zhang2026searchgymbootstrappingrealworldsearch,
      title={SearchGym: Bootstrapping Real-World Search Agents via Cost-Effective and High-Fidelity Environment Simulation}, 
      author={Xichen Zhang and Ziyi He and Yinghao Zhu and Sitong Wu and Shaozuo Yu and Meng Chu and Wenhu Zhang and Haoru Tan and Jiaya Jia},
      year={2026},
      eprint={2601.14615},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Acknowledgments

This project is built upon the outstanding work of:

• AReaL - A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning and Agents, developed by the AReaL Team at Ant Group and Tsinghua IIIS.
• ASearcher - An Open-Source Large-Scale Reinforcement Learning Project for Search Agents.

We are deeply grateful to the authors and contributors of these projects for their pioneering work in asynchronous RL training and search agent development.