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

Abstract: Training deep research agents requires long-horizon trajectories that interleave search, evidence aggregation, and multi-step reasoning. However, existing data collection pipelines typically rely on proprietary web APIs, making large-scale trajectory synthesis costly, unstable, and difficult to reproduce. We present OpenResearcher, a reproducible pipeline that decouples one-time corpus bootstrapping from multi-turn trajectory synthesis and executes the search-and-browse loop entirely offline using three explicit browser primitives: search, open, and find, over a 15M-document corpus. Using GPT-OSS-120B as the teacher model, we synthesize over 97K trajectories, including a substantial long-horizon tail with 100+ tool calls. Supervised fine-tuning a 30B-A3B backbone on these trajectories achieves 54.8\% accuracy on BrowseComp-Plus, a +34.0 point improvement over the base model, while remaining competitive on BrowseComp, GAIA, and xbench-DeepSearch. Because the environment is offline and fully instrumented, it also enables controlled analysis, where our study reveals practical insights into deep research pipeline design, including data filtering strategies, agent configuration choices, and how retrieval success relates to final answer accuracy. We release the pipeline, synthesized trajectories, model checkpoints, and the offline search environment at https://github.com/TIGER-AI-Lab/OpenResearcher.

• The paper presents a fully open offline pipeline that synthesizes long-horizon research trajectories using explicit browser primitives.
• It achieves a 34-point accuracy improvement over baseline models by integrating gold document bootstrapping and multi-step reasoning without online API dependencies.
• The method enables reproducible, cost-effective corpus construction and detailed analysis, laying groundwork for advanced deep research agent development.

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

Introduction

OpenResearcher addresses the central challenge of generating high-quality, long-horizon research trajectories for training and evaluating deep research agents. Existing agentic systems designed for iterative search, evidence aggregation, and multi-step reasoning have been critically limited by cost, instability, and reproducibility constraints due to their reliance on live, proprietary web APIs. OpenResearcher proposes a shift to a fully open, offline trajectory synthesis pipeline, introducing a scalable and transparent method built on three explicit browser primitives—search, open, and find. The system enables both cost-effective corpus construction and detailed, instrumented analysis, offering a practical foundation for the next generation of deep research agents.

Figure 1: Overview of the OpenResearcher trajectory synthesis pipeline, highlighting curation of long-horizon queries, one-time gold document bootstrapping to build a 15M-document offline corpus, and trajectory generation via a teacher model using search, open, and find tools.

Pipeline Architecture and Offline Synthesis

The key methodological innovation is a decoupled, offline synthesis pipeline. Construction begins with challenging multi-hop questions curated from MiroVerse, selected to require sustained and interleaved reasoning beyond trivial retrieval. For each question, a one-time online bootstrapping phase retrieves gold documents known to support the answer, which are merged with 15 million distractor documents from FineWeb, producing an offline corpus optimized for both realistic coverage and controlled experimentation.

The environment exposes three browser primitives:

• Search: Dense retrieval of top-$K$ snippets for a query using an FAISS index built on Qwen3-Embedding-8B representations;
• Open: Access to full document content;
• Find: In-document string matching for rapid evidence localization.

By instrumenting a teacher LLM (GPT-OSS-120B) with these tools, the pipeline synthesizes over 97K trajectories, capturing diverse reasoning paths—including a significant distributional tail requiring 100+ tool calls.

Figure 2: OpenResearcher's browser abstraction exposes search, open, and find, enabling multi-scale, explicit navigation from corpus to passage to fine-grained evidence.

Empirical Evaluation and Analysis

Main Results

Fine-tuning a 30B-A3B backbone on the resultant dataset yields marked improvements on BrowseComp-Plus, the current standard for benchmarked deep research agents. The accuracy attained is 54.8%, outperforming GPT-4.1 (36.4%), Claude-4-Opus (36.8%), and DeepSeek-R1 (16.4%)—representing a +34.0 point increase over the base model, and demonstrating that high-fidelity, offline-synthesized trajectories alone induce substantial gains without requiring reinforcement learning or additional online interactions.

Generalization to live-web benchmarks, including BrowseComp, GAIA, and xbench-DeepSearch, is robust: OpenResearcher achieves 26.3%, 64.1%, and 65.0% on these tasks, rivaling proprietary and larger-scale open-source systems. This is achieved without any further fine-tuning on live-web data, a direct testament to the transferability of the pipeline's synthesized signals.

Trajectory Distributions and Tool Usage

The synthesized dataset reveals clear distinctions in search efficiency: failed trajectories average nearly twice as many tool calls (71.7 vs 38.4), dominated by excessive search rather than inefficient navigation or evidence localization. This strongly suggests that suboptimal search formulation, rather than mere lack of exploration, is the typical cause of failure in long-horizon research.

Figure 3: Distribution of tool calls by outcome and analysis of tool usage; failures arise from repeated search attempts, not document navigation.

The Pass@$k$ analysis indicates that increasing the number of sampled trajectories per question raises the overall solve rate—Pass@1 of 0.567 increases to 0.792 at Pass@16—suggesting considerable solution diversity and confirming the necessity of broad sampling to cover reasoning path variability.

Ablations and Empirical Findings

Several strong claims emerge from ablation studies:

• Final-answer correctness is not necessary as a filtering signal for trajectory inclusion in SFT; both correct and incorrect traces impart similar downstream improvements, underlining the value of learning from structured exploration patterns irrespective of answer outcome.
• One-time gold document bootstrapping is essential: removing gold documents reduces gold-document hit rates (29.54% $\rightarrow$ 1.73%) and drastically collapses benchmark accuracy (54.81% $\rightarrow$ 6.35%), highlighting the non-negotiable role of corpus coverage.
• Longer turn budgets yield diminishing returns beyond 100 turns; initial increases directly improve both answer accuracy and gold-document exposure, but the effect saturates, defining an empirical horizon for agentic exploration sufficiency.

Explicit browser primitives play a critical role:

• Restricting to search-only dramatically underperforms (43.86%), with near-zero gold document findings. Adding open and find increases accuracy to 62.17% and robustly ensures evidence localization and efficient reasoning chains.

Evidence Exposure and Reasoning

Figure 4: Timing and coverage of gold-document exposure in trajectories on BrowseComp-Plus, showing final-answer accuracy conditional on evidence access.

Mere retrieval of supporting documents does not guarantee correct answers. The probability of correctness given a search hit is 61.84%, but rises to 86.72% when the evidence is explicitly opened and found. Correct trajectories nearly always involve at least one gold-document open (95.01%), underscoring that successful deep research depends conjointly on retrieval and downstream reasoning over identified content.

Cost Efficiency and Scalability

OpenResearcher's offline design eliminates API cost and rate limits, reducing the total price of trajectory synthesis by orders of magnitude compared to commercial web search providers. It further enables perfect reproducibility and parallelism, making it uniquely suited for large-scale, community-driven agentic development and systematic failure analysis.

Implications and Future Directions

Practically, OpenResearcher provides a scalable and reproducible platform for trajectory-based research agent training without dependence on proprietary or unstable external APIs. The explicit offline environment and open-source release structure lower the barrier for further research on agentic tool use, trajectory synthesis optimization, and experimental ablation. The pipeline's modularity allows new question sets, corpus augmentations, or alternative tool designs to be introduced systematically.

Theoretically, the empirical analyses clarify that supervision signals for research agents extend beyond answer correctness, shifting focus to the structure of information-seeking processes and error modes in sequence-level search strategies. The distinction between search, open, and find reveals discrete bottlenecks in existing architectures that can be more directly targeted with future reinforcement learning, curriculum design, or dynamic tool instantiation.

Prospective developments may leverage this controlled environment for interactive RL, meta-agency, or curriculum learning in complex, multi-agent settings. The pipeline's detailed instrumentation will support programmatic diagnosis of failure types, and its open dataset will likely serve as a benchmark for progress in long-horizon research reasoning, interpretability, and generalization.

Conclusion

OpenResearcher operationalizes a fully open, offline deep research agent synthesis pipeline capable of rivaling proprietary models on rigorous benchmarks. By fundamentally decoupling corpus bootstrapping from trajectory generation, and exposing explicit, multi-scale browsing tools, it realizes reproducible, instrumented, and analytically tractable large-scale agent training. The work establishes that high-quality synthetic trajectories alone can suffice for significant downstream gains and lays important groundwork for ongoing research on web-scale reasoning, agent supervision, and the design of practical generalist research systems.

Figure 5: Pass@$k$ progression for $k \in \{1,2,4,8,16\}$, illustrating the relationship between sampling budget and trajectory solve rate across diverse queries.