QUEST: Training Frontier Deep Research Agents with Fully Synthetic Tasks

Deep research agents extend the role of search engines from retrieving keyword-matched pages to synthesizing knowledge, fundamentally changing how humans interact with information. However, frontier systems remain proprietary, while existing open agents often generalize poorly across different task types, leaving unclear how to train a broadly capable deep research agent. We release QUEST, a family of open models (ranging from 2B to 35B) that serve as general-purpose deep research agents designed to handle a wide range of long-horizon search tasks, with strong capabilities in fact seeking, citation grounding, and report synthesis. To build QUEST, we propose an effective training recipe combining mid-training, supervised fine-tuning, and reinforcement learning. Central to this recipe is a curated data synthesis pipeline based on unified rubric trees, which applies to different task types and enables synthesizing training data with verifiable rewards without human annotation. In addition, QUEST incorporates a built-in context management mechanism that enables effective long-horizon reasoning and knowledge synthesis. Using only 8K synthesized tasks, QUEST approaches or even surpasses frontier closed-source agents across eight deep research benchmarks spanning diverse task types, and achieves the best overall performance among recent open-weight agents. We released everything: models, data, and training scripts.

📋 한 줄 요약

**[Deep Research Agent / 합성 데이터]** QUEST가 2B~35B open model family를 8K 합성 task만으로 학습 — unified rubric tree로 task type 무관 합성 가능, 8 deep research 벤치에서 frontier closed agent에 근접·능가, open-weight SOTA.

🎯 핵심 기여도

Deep research agent가 검색 엔진을 keyword 매칭에서 knowledge synthesis로 확장해 human-information 상호작용을 근본 변혁하지만, frontier system은 proprietary이고 기존 open agent는 task type 간 일반화가 부족한 한계 진단.
QUEST 공개 — 2B~35B 범위 open model family로 fact seeking·citation grounding·report synthesis 강점을 가진 general-purpose deep research agent.
효과적 training recipe 제안 — mid-training + supervised fine-tuning + reinforcement learning 조합.
핵심은 unified rubric tree 기반 curated data synthesis pipeline — task type 무관하게 적용 가능, human annotation 없이 verifiable reward 데이터 합성. Built-in context management로 long-horizon reasoning·knowledge synthesis 지원.

💡 핵심 아이디어

Deep research agent의 broad 능력 달성은 task별 데이터 큐레이션이 아니라 unified rubric tree로 task type 무관 합성 가능한 verifiable-reward 데이터 + mid-training/SFT/RL의 체계 레시피 + long-horizon context management의 결합으로 가능하며, 단 8K 합성 task만으로 frontier closed agent에 근접·능가할 수 있다.

🔬 기술적 접근법

**방법론**: QUEST — Mid-training + SFT + RL + unified rubric tree 합성.
**핵심 기법**: (1) 2B~35B 범위 open model family, (2) Mid-training → SFT → RL의 3단계 학습, (3) Unified rubric tree 기반 data synthesis pipeline으로 task type 무관 합성, (4) Verifiable reward로 human annotation 없이 학습 데이터 확보, (5) Built-in context management로 long-horizon reasoning·knowledge synthesis.

📊 주요 결과

단 8K synthesized task만으로 학습.
8 deep research 벤치마크에서 frontier closed-source agent에 근접 또는 능가.
최근 open-weight agent 중 최고 종합 성능.
모델·데이터·학습 스크립트 전부 공개.

💭 의의 및 한계

**의의**: Frontier deep research agent의 open 대안을 제공, unified rubric tree로 task 합성의 일반 레시피 정립, 8K task의 데이터 효율성, 모델·데이터·코드 전체 공개로 community 가속. **한계**: 8K 합성 task의 다양성·realistic distribution shift 처리는 후속, rubric tree 설계 품질이 verifiable reward 효과를 결정, very long-horizon task에서의 context management 한계 가능.

🚀 실용적 활용

Deep research agent open-source 표준.
Synthetic data 기반 agent 학습 일반 레시피.
Fact-seeking·citation·report 자동 합성 응용.