README.md

April 21, 2026 · View on GitHub

A Scalable Agentic RL Training Framework for Deep Research Agent

Paper · Project Page · Model

LiteResearcher is a training framework that makes Agentic RL scalable for deep research agents. By constructing a lite virtual world that mirrors real-world search dynamics, we enable a continuously improving training recipe that empowers a tiny 4B search agent to outperform large-scale open-source and commercial models.

LiteResearcher-4B achieves 71.3% on GAIA and 78.0% on Xbench-DeepSearch, surpassing models up to 8× larger (Tongyi DeepSearch 30B, WebSailor 30B) and matching commercial systems (Claude-4.5-Sonnet, GPT-5).

Results

ModelGAIA-TextBrowseCompBrowse.(ZH)HLEFramesWebWalkerSeal-0Xbench-DS
Commercial Models
Claude-4-Sonnet68.3%12.2%29.1%20.3%80.7%61.7%64.6%
Claude-4.5-Sonnet71.2%19.6%40.8%24.5%85.0%53.4%66.0%
DeepSeek-V3.263.5%67.6%65.0%40.8%80.2%38.5%71.0%
DeepSeek-V3.163.1%30.0%49.2%29.8%83.7%61.2%71.0%
Minimax-M275.7%44.0%48.5%31.8%72.0%
OpenAI-GPT-5-high76.4%54.9%65.0%35.2%51.4%77.8%
GLM-4.671.9%45.1%49.5%30.4%70.0%
Kimi-Researcher26.9%78.8%36.0%69.0%
Kimi-K2-090560.2%7.4%22.2%21.7%58.1%25.2%61.0%
Open-Source Models
Mirothinker 8B66.4%31.1%40.2%21.5%80.6%60.6%40.4%60.6%
Tongyi Deepsearch 30B70.9%43.4%46.7%32.9%90.6%72.2%75.0%
ASearcher QWQ v2 32B58.7%74.5%51.1%
WebSailor 30B53.2%53.3%
WebDancer 32B (QwQ)51.5%3.8%18.0%47.9%38.3%
WebExplorer 8B50.0%15.7%32.0%17.3%75.7%62.7%53.7%
DeepMiner 32B58.7%33.5%40.1%62.0%
AFM-RL 32B55.3%11.1%18.0%63.0%
SFR-DeepResearch 20B66.0%28.7%82.8%
AgentCPM-Explore 4B63.9%24.1%29.1%19.1%82.7%68.1%40.5%70.0%
LiteResearcher-4B71.3%27.5%*32.5%*22.0%83.1%72.7%41.8%78.0%

Best open-source results in bold. Results with * use a 64k context window with a memory mechanism.

Method Overview

Three pillars enable scalable Agentic RL:

  1. Co-construct Training Data & Corpus — Scale up information sources with a simple-but-effective synthesis pipeline, then co-evolve training QA pairs and the local webpage corpus.
  2. Stable Local Tool Environment — Build local search engine (Milvus + BGE-M3) and local browse tool (PostgreSQL) from ~32M real webpages, achieving 10–46× speedup at zero marginal cost.
  3. Difficulty-Aware Curriculum RL — Multi-stage curriculum with on-policy GRPO, filtering tasks by pass@8 difficulty to sustain monotonic improvement.

Repository Structure

├── Inference/              # Inference & evaluation (released)
├── Training/               # RL training (coming soon)
├── DataGen/                # Data synthesis (coming soon)
├── Environment/            # Local search/browse environment (coming soon)
└── docs/                   # Project page

Quick Start — Evaluation

cd Inference
pip install -r requirements.txt
cp .env.example .env
# Edit .env: set MODEL, SERPER_KEY_ID, SCRAPEDO_API_KEY

# Start model server (SGLang/vLLM)
bash scripts/start_sglang.sh

# Run evaluation
bash scripts/run_all.sh

See Inference/README.md for detailed configuration and usage.

Release Plan

  • Evaluation code
  • Project page
  • Model weights (LiteResearcher-4B)
  • Training code (GRPO + curriculum RL)
  • Data synthesis pipeline
  • Local search/browse environment setup

Citation

@article{li2026literesearcher,
  title={LiteResearcher: A Scalable Agentic RL Training Framework for Deep Research Agent},
  author={Wanli Li and Bince Qu and Bo Pan and Jianyu Zhang and Zheng Liu and Pan Zhang and Wei Chen and Bo Zhang},
  journal={arXiv preprint arXiv:2604.17931},
  year={2026}
}

License

Apache 2.0