TODO:

• generate trajectory details tools?

Blog Release Xueguang

Zhuofeng Li*$^{1}$, Dongfu Jiang*$^{2}$, Xueguang Ma*$^{2}$, Haoxiang Zhang$^{3}$, Yuyu Zhang$^{4}$, Kai Zou$^{5}$, Ping Nie$^{2}$, Jianwen Xie$^{6}$, Yu Zhang†$^{1}$, Wenhu Chen†$^{2}$

$^{1}$Texas A&M University $^{2}$Waterloo University $^{3}$UC San Diego $^{4}$Verdent AI $^{5}$NetMind AI $^{6}$Lambda

*****: Equal Contribution; **†**Corresponding Authors

February 2026

<aside> 🌟

TL;DR

• Using GPT-OSS-120B + offline corpus and retriever, you can synthesize long-horizon deep research trajectories with 100+ turns with no additional cost for search and scraper APIs at training time.
• SFT from deep research trajectory significantly boosts performance: training Nemotron-3-Nano-30B-A3B-Base on our synthesized trajectories improves BrowseComp-Plus accuracy to 54.8%.
• Offline corpus eliminates API dependency: API-based synthesis is expensive ($0.001-0.01/query), rate-limited, and non-deterministic. Our offline setup is free, unlimited, and fully reproducible.
• Process over outcome: even trajectories without correct final answers teach models how to search effectively. TODO: delete here
• We release everything: search environment, deep research trajectories, trained models, and code.

👨‍💻 Github, 🤗 HF Model, 🤗 HF Dataset, 📈 Wandb Logs, 🔎 Eval Logs

</aside>

1. Open Source Gaps in Deep Research Agents

Deep research agents—systems that perform iterative search, evidence gathering, and multi-step reasoning—have become a key frontier for LLM capabilities. Commercial systems like Perplexity Deep Research, OpenAI Deep Research, and Gemini Deep Research demonstrate impressive performance, but open-source alternatives lack key components.

Key gaps in existing work:

• No long-horizon trajectories: Search-R1 uses Wikipedia with only 2-5 turns—far from real deep research
• No offline environment: Most rely on live search APIs, making reproduction expensive and non-deterministic
• Incomplete releases: No single work releases weights + trajectories + code + environment together

This post provides a fully open pipeline for synthesizing long-horizon deep research trajectories with 100+ turns, releasing all components for the community.