OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis

arXiv 2026 52.4 method

TLDR

OpenResearcher is an open pipeline for synthesizing long-horizon deep research trajectories offline, achieving strong results on research benchmarks.

Reasoning

The paper presents a fully open, reproducible pipeline for generating deep research trajectories, with strong empirical results on multiple benchmarks. However, it relies on a teacher model and an offline corpus, which may limit generalization to live web environments.

Read-first score

Read-first score 52.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 55.

Recency 8%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=analysis

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=checkpoint,github,reproduce

Topical relevance 42%

45.8

Uses existing LLM keyword relevance scores normalized to 0-100. AI scientist,automated scientific discovery,autonomous research agent,automated research,literature review agent,survey generation,automated experimentation,experiment design agent,AI for scientific research,paper writing agent,research automation,scientific discovery agent

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 27.

Keyword Scores

automated research

research automation

autonomous research agent

scientific discovery agent

AI for scientific research

automated scientific discovery

literature review agent

AI scientist

survey generation

automated experimentation

experiment design agent

paper writing agent

Deep Analysis

Innovations

Fully open, reproducible pipeline for long-horizon deep research trajectory synthesis that decouples one-time corpus bootstrapping from multi-turn trajectory synthesis
Offline search-and-browse loop using three explicit browser primitives (search, open, find) over a 15M-document corpus, eliminating reliance on proprietary web APIs
Synthesis of over 97K trajectories, including a long-horizon tail with 100+ tool calls, using GPT-OSS-120B as the teacher model
Supervised fine-tuning of a 30B-A3B backbone on the synthesized trajectories, achieving 54.8% on BrowseComp-Plus (+34.0 points) and competitive performance on BrowseComp, GAIA, and xbench-DeepSearch
Release of the entire pipeline, synthesized trajectories, model checkpoints, and offline search environment

Methodology

OpenResearcher decouples one-time corpus bootstrapping from multi-turn trajectory synthesis, executing an offline search-and-browse loop with three browser primitives (search, open, find) over a 15M-document corpus. A GPT-OSS-120B teacher model synthesizes long-horizon trajectories, which are then used to fine-tune a 30B-A3B backbone via supervised learning.

Key Results

The fine-tuned model achieves 54.8% accuracy on BrowseComp-Plus, a +34.0 point improvement over the base model, and remains competitive on BrowseComp, GAIA, and xbench-DeepSearch.

Resources

Paper