Research map / Multi-agent systems and games / LLM agents

Quest: Training Frontier Deep Research Agents with Fully Synthetic Tasks

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces Quest, open deep research agents (2B to 35B) trained on fully synthetic rubric-tree tasks with a context condenser, mid-training, SFT, and rubric/fact-checking RL.

Problem

Open deep research agents generalize poorly across task types (fact seeking, report synthesis, citation grounding), and existing synthetic training data mostly uses single verifiable answers with binary rewards.

Approach

Synthesize tasks paired with fully synthetic hierarchical rubric trees and scripted evaluators, use a structured Context Condenser (trusted/untrusted/uncertain state) for long-horizon context management, and train with mid-training, SFT, and GRPO using rubric-based plus citation fact-checking rewards.

Key results

• Quest-35B reaches 64.6% on BrowseComp with discard-all context management (vs. 67.8 Claude Opus 4.5 and 59.9 GPT-5), 30.7% on Mind2Web 2 (vs. 28.0 OpenAI-DR), and 48.2% on DeepResearch Bench (vs. 47.0 OpenAI-DR).
• Quest-35B achieves 80.8% on GAIA, above GPT-5's 76.4%, and is best among open-weight agents on most benchmarks.
• Even Quest-2B-SFT achieves competitive HLE and GAIA scores, matching or outperforming o3 on those benchmarks (o3 at 24.9% HLE, 70.5% GAIA as reported).

Surprising findings

• Simple SFT degrades open-ended report performance relative to the vanilla base model, while RL substantially improves open-ended tasks.
• Search result prediction mid-training objective hurt overall performance when combined with the other mid-training objectives.

Limitations

• Small models (2B) lag far behind on report synthesis tasks such as DeepResearch Bench and LiveResearchBench.
• Generated evaluation scripts can contain errors (redundant nodes, literal-content verification, missing negative judgments), found in 6 of 50 sampled tasks.
• Vanilla Qwen3.5 results could not be reproduced against the vendor's reported numbers, and RL slightly hurts HLE and GAIA.

Experimental setup

Datasets: Quest-8K synthetic tasks Google Trends keyword seeds Visit cache HTML extraction triplets

Benchmarks: BrowseComp BrowseComp-Plus Mind2Web 2 WideSearch Humanity's Last Exam GAIA DeepResearch Bench LiveResearchBench

Metrics: accuracy rubric-based score pairwise score relative to reference report

Baselines: OpenAI DeepResearch GPT-5 Claude Opus 4.5 Gemini 3 Pro Tongyi DeepResearch OpenResearcher DR Tulu o3

Builds on: GRPO Mind2Web 2 rubric tree DeepResearch Bench rubric criteria Tongyi DeepResearch VERL

Closest papers by meaning

• OpenSearch-VL: An Open Recipe for Frontier Multimodal Search Agents (NeurIPS 2026 · similarity 0.89)
• Fathom-DeepResearch: Unlocking Long Horizon Information Retrieval and Synthesis for SLMs (ICLR 2026 · similarity 0.89)
• Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization (NeurIPS 2026 · similarity 0.89)
• Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.89)
• HierSearch: A Hierarchical Enterprise Deep Search Framework Integrating Local and Web Searches (AAAI 2026 · similarity 0.88)
• WebWatcher: Breaking New Frontiers of Vision-Language Deep Research Agent (ICLR 2026 · similarity 0.88)