LiteResearcher

A Low-Cost, Scalable Agentic RL Training Framework for Deep Research Agent

If you like our project, please give us a star ⭐ on GitHub for the latest update.

News

2026-06

• 🚀 Training code released — GRPO + difficulty-aware curriculum (training/)
• 📊 Training data released — Stage-1 & Stage-2 prompts (LiteResearcher-Data)
• 🧊 SFT cold-start checkpoint released (LiteResearcher-4B-SFT)
• 🛠️ Data synthesis pipeline released (datagen/)
• 🌐 Local search/browse environment released (environment/)
• 📚 32M-record search corpus released (LiteResearcher-Corpus)

2026-04

• 🎯 RL model weights released (LiteResearcher-4B)
• 📈 Evaluation code & project page released

LiteResearcher-4B is a 4B deep research agent that matches frontier systems at a fraction of the size — trained with $0 marginal API cost by replacing live-web interaction with a stable local search/browse environment that mirrors real-world search dynamics.

Highlights

• Open-source SOTA — 71.3% GAIA / 78.0% Xbench-DS, beating 30B open-source agents and surpassing Claude-4.5-Sonnet on GAIA and GPT-5-high on Xbench-DS.
• +15.7 GAIA points from RL — SFT 55.6% → RL 71.3%, vs. only +3.8 for AgentCPM-Explore when training with live web interaction.
• $0 marginal API cost — 73.2M local tool calls during RL; the same volume would cost $59K–$243K via live search/browse APIs.

Left: Xbench-DeepSearch accuracy vs. model size — our 4B model reaches 78.0%, matching/surpassing 100×+ larger systems. Right: Average rollout time and cost per turn — LiteResearcher is the fastest and cheapest.

Results

Comparison across commercial models and open-source deep research agents on eight benchmarks. Best score among ≤8B models is in bold; LiteResearcher-4B leads on 6 of 8.

Across all 8 benchmarks, LiteResearcher-4B is the best ≤8B agent on 6 — Mirothinker-8B leads on BrowseComp and BrowseComp-ZH. Full numbers are also in the paper and training/README.md.

Method Overview

Three pillars enable low-cost, 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, enabling the RL stage to run fully locally with no API consumption, 10–46× speedup, and zero marginal tool cost.
3. Difficulty-Aware Curriculum RL — Multi-stage curriculum with on-policy GRPO, filtering tasks by pass@8 difficulty to sustain monotonic improvement.

Trajectory Cases

We release 15 hand-audited rollout trajectories from LiteResearcher-4B across 8 deep-research benchmarks (GAIA, Xbench-DS, Frames, HLE, Seal-0, WebWalker, BrowseComp, BrowseComp-ZH).

🔎 Live viewer: https://simplexai-labs.github.io/LiteResearcher/cases/

Each trajectory renders 40–170 steps showing the model's think → search → visit → answer chain, with tool queries, visited URLs, and tool responses inline. Source data lives under docs/cases/.

Repository Structure

├── inference/              # Inference & evaluation (released)
├── training/               # RL training — GRPO + curriculum (released)
├── datagen/                # Data synthesis (released)
├── environment/            # Local search/browse environment (released)
└── docs/                   # Project page

Quick Start — Evaluation

cd inference
pip install -r requirements.txt
cp .env.example .env
# Edit .env: set MODEL, SERPER_KEY_ID (browser uses Jina Reader by default; set SCRAPEDO_API_KEY only if using BROWSER_PROVIDER=scrapedo)

# 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.

Quick Start — Training

The full two-stage RL training pipeline (GRPO + TIS + difficulty-aware curriculum) is in training/, and the training data is hosted on 🤗 LiteResearcher-Data.

Prerequisites

• GPU — Stage 1: 8×H20 (1 node); Stage 2: 16×H20 (2 nodes).
• Local tool backend — RL runs against the local search/browse environment, not live web. Bring up the search service (Milvus + Redis) and the browse service (PostgreSQL) before training. See environment/ for the search backend and examples/sglang_multiturn/search_browser/tool_backend/ for the browse backend.

1. Install

cd training
pip install -e .[sglang]                   # verl-based training stack

2. Configure the tool backend

cp examples/sglang_multiturn/search_browser/tool_backend/.env.example \
   examples/sglang_multiturn/search_browser/tool_backend/.env
# Edit .env: PG_* (browse DB), SUMMARY_API_*, LLM_JUDGE_API_*, optional SCRAPEDO_API_KEY

# Start the browse service (reads the .env above)
bash examples/sglang_multiturn/search_browser/tool_backend/start_browse.sh

3. Download the training data

hf download simplex-ai-inc/LiteResearcher-Data --repo-type dataset \
            --local-dir ./literesearcher_data    # 28K prompts, 19 MB

4. Stage 1 — RAG-only warmup (8×H20, 32K ctx)

export TRAIN_DATA=./literesearcher_data/stage1/train.parquet
export VAL_DATA="$TRAIN_DATA"     # no separate val bundled; verl needs a non-empty val_files
export MODEL_PATH=$(hf download simplex-ai-inc/LiteResearcher-4B-SFT \
                                --local-dir ./literesearcher_sft)
bash examples/sglang_multiturn/search_browser/stage1_rag_only.sh

5. Stage 2 — mixed curriculum (16×H20, 48K ctx)

Resume from a Stage-1 checkpoint (around step 220).

export TRAIN_DATA=./literesearcher_data/stage2/train.parquet
export VAL_DATA="$TRAIN_DATA"
export MODEL_PATH=/path/to/stage1-ckpt/global_step_220
bash examples/sglang_multiturn/search_browser/stage_2_mix_rag_on_policy_48k.sh

See training/README.md for the full reproduction recipe (including the SFT cold-start prerequisite, environment variables, and config knobs) and the dataset card for the data schema and curriculum design.

Release Plan

• Evaluation code
• Project page
• Model weights — RL (LiteResearcher-4B)
• Model weights — SFT cold-start (LiteResearcher-4B-SFT, built on Qwen3-4B-Thinking-2507) 🆕
• Local search/browse environment setup (environment/)
• Search corpus — 32M records (LiteResearcher-Corpus)
• Training code — GRPO + curriculum RL (training/)
• Training data — Stage-1 & Stage-2 prompts (LiteResearcher-Data)
• Data synthesis pipeline (datagen/)

Acknowledgements

LiteResearcher's training stack is built on verl, ByteDance's RL training library, which we fork and extend with the multi-turn search/browse agent loop, difficulty-aware curriculum, and local-tool reward pipeline. We also build on SGLang for rollout serving, Qwen3 as the base model, and Milvus + BGE-M3 for the local search environment. We thank these projects and their communities.

Contributing

Contributions are welcome — see CONTRIBUTING.md for development setup, pull-request guidelines, and our Code of Conduct.

Powered By

LiteResearcher is the engine behind lev8, Simplex AI's parallel agentic search platform — frontier-grade deep research, fast and cheap enough to run hundreds of agents per query. Explore → lev8.com

Citation

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

License

Released under the Apache License 2.0.