PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Zhuofeng Li, Dongfu Jiang, Xueguang Ma, Haoxiang Zhang, Ping Nie, Yuyu Zhang, Kai Zou, Jianwen Xie, Yu Zhang, Wenhu Chen

arXiv:2603.20278v1cs.IRcs.AIcs.CL

TL;DR

Existing deep-research trajectory pipelines are difficult to scale, reproduce, and analyze because they depend on scarce trajectories and unstable live-web search. OpenResearcher moves synthesis offline after one-time corpus bootstrapping, uses explicit browser operations, and shows that corpus coverage and browser interaction choices strongly affect downstream accuracy.

  • Problem

    Existing approaches lack a scalable, low-cost way to generate realistic long-horizon deep-research trajectories, while live-web APIs make synthesis expensive, unstable, and difficult to analyze reproducibly.

  • Method

    OpenResearcher performs one-time online bootstrapping for corpus coverage, then synthesizes trajectories offline using explicit search, open, and find browser operations.

  • Results

    54.81% BrowseComp-Plus accuracy falls to 6.35% without bootstrapped gold documents, while adding open and find raises teacher-side accuracy to 62.17%.

  • Takeaways & Limitations

    Effective offline trajectory synthesis depends on corpus coverage and explicit evidence-access operations, while longer horizons help until performance plateaus beyond roughly 100 turns.

  • Takeaways & Limitations

    Retrieving the gold document does not guarantee correctness because reasoning errors, such as misreading table columns, can still produce wrong answers.

Abstract

from arXiv · show

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.

1 Introduction

OpenResearcher addresses the cost, instability, and limited controllability of live-web trajectory synthesis by moving generation offline after one-time corpus bootstrapping. Its open pipeline uses explicit browsing primitives and produces long-horizon trajectories that support strong deep-research performance and controlled analysis.

  • Motivation: Existing deep-research trajectories are scarce, while live-web synthesis is expensive, unstable, and difficult to analyze reproducibly.These limitations arise because failed search paths incur API costs, live-web content changes over time, and internal search events are difficult to control.
  • Approach: OpenResearcher decouples one-time online corpus bootstrapping from fully offline multi-turn trajectory generation.The pipeline seeds answer-supporting documents, builds an offline corpus and search engine, and then runs synthesis locally.
  • Approach: Its minimal browser abstraction exposes search, open, and find so the teacher model learns retrieval, document inspection, and evidence localization.The explicit operations support systematic information seeking and multi-scale knowledge discovery.
  • Results: Over 97K trajectories are synthesized over a 15M-document corpus, including a substantial tail requiring 100+ tool calls.GPT-OSS-120B serves as the teacher model for this long-horizon synthesis.
  • Analysis: The offline environment is cheaper, reproducible, and instrumented for controlled studies of filtering, corpus construction, agent configuration, and retrieval success.The fixed corpus, search backend, and browser actions make internal search events traceable beyond benchmark accuracy.
  • Contribution: OpenResearcher is presented as a fully open-source pipeline whose resulting model rivals proprietary systems on long-horizon search and reasoning tasks.The authors position the released tools, trajectories, and analyses as resources for studying search supervision and future agent design.

2 Preliminary

Deep research agents repeatedly interleave reasoning, tool actions, and environmental observations until producing a final answer. The resulting trajectory records the full history that conditions each subsequent policy decision.

  • Workflow: A deep-research trajectory begins with a query, system prompt, and tool metadata, followed by reasoning–action–observation triplets.The trajectory ends with a final answer action.
  • Workflow: At step t, the policy generates reasoning r_t and action a_t from the history H_t−1 of previous interactions.The environment then executes the action and returns an observation.
  • Workflow: The environment maps each tool action a_t to an observation o_t and updates the trajectory with the resulting interaction.This creates the sequential state used by later reasoning and actions.
  • Termination: The loop continues until the model stops issuing tool calls and outputs the final answer a_T.This iterative process grounds reasoning in external evidence.

3 Offline Trajectory Synthesis

OPENRESEARCHER makes long-horizon trajectory synthesis reproducible by separating one-time corpus bootstrapping from offline generation and exposing browsing through explicit search, open, and find operations. It uses these components to produce filtered trajectories for training smaller reasoning models.

  • Question and trajectory preparation: The pipeline collects challenging long-horizon questions and regenerates trajectories from clean question–answer pairs rather than using partial source traces.The selected questions require multi-hop reasoning over heterogeneous evidence, while existing partial trajectories may contain unsupported evidence or degenerate tool patterns.
  • Offline corpus construction: One-time answer-guided bootstrapping retrieves gold documents before synthesis, reducing ambiguity between search failure and missing corpus evidence.The process uses question–answer queries, cleans and deduplicates retrieved content, and extracts 10K gold documents for 6K questions.
  • Offline corpus construction: The offline corpus merges answer-supporting gold documents with 15 million FineWeb distractor documents and indexes them for dense retrieval with Qwen3-Embedding-8B and FAISS.At inference time, natural-language queries return ranked documents through the locally served search engine.
  • From search to real browsing: The browsing abstraction exposes search for candidate sources, open for full documents, and find for exact evidence localization across progressively narrower information scales.This explicit structure supports evidence discovery, verification, and synthesis while more closely reflecting real browsing than search-only retrieval.
  • Trajectory generation procedure: GPT-OSS-120B generates trajectories using only the three tools and stepwise reasoning, after which filtering removes overlong, malformed, or inconclusive traces.The resulting dataset contains 97K+ trajectories, including many requiring more than 100 tool calls, for supervised fine-tuning of smaller reasoning models.

4 Experiments

Experiments evaluate the synthesized trajectories, their downstream training value, and design choices affecting offline deep research. Results show strong benchmark performance, benefits from explicit browser tools and corpus bootstrapping, and distinct patterns linking search behavior to correctness.

  • Main Results: 54.8% accuracy on BrowseComp-Plus substantially exceeds GPT-4.1 at 36.4%, Claude-4-Opus at 36.8%, and the base model at 20.8%.The improvement over the base Nemotron-3-Nano-30B-A3B model is +34.0 absolute points.
  • Main Results: 26.3%, 64.1%, and 65.0% accuracy on BrowseComp, GAIA, and xbench-DeepSearch respectively remains competitive with frontier models and exceeds listed open-source systems.The evaluation includes both closed-web and live-web deep research benchmarks.
  • Trajectory Analysis: Failed trajectories average 71.7 tool calls versus 38.4 for correct trajectories, with search accounting for most of the excess.Search usage is 48.7 versus 22.1 calls, while find usage remains similar at 3.2 versus 2.8.
  • Trajectory Analysis: Pass@k rises from 0.567 at Pass@1 to 0.792 at Pass@16, while solve rates are bimodal across questions.Over 20% of questions have pass rates near 0%, approximately 30% reach near 100%, and the remainder occupy intermediate rates.
  • Ablations: Removing bootstrapped gold documents lowers gold-document hit rate from 29.54% to 1.73% and BrowseComp-Plus accuracy from 54.81% to 6.35%.The ablation supports one-time online bootstrapping as essential for effective offline corpus coverage and downstream post-training.
  • Ablations: Adding open to search raises accuracy from 43.86% to 56.39%, while adding find reaches 62.17% and improves evidence access.With search, open, and find, gold-document hit rate reaches 53.37% and the first gold hit moves from 20.60 to 17.23 turns.

5 Related Work

Deep research agents extend prior tool-use systems by sustaining iterative search, evidence aggregation, and multi-step reasoning over long horizons. Offline and synthetic environments address the cost, rate-limit, and reproducibility challenges of live-web interaction.

  • Deep research agents must sustain exploration across many tool calls, reconcile heterogeneous sources, and decide when enough evidence has been gathered.
  • Synthetic and offline environments provide reproducible testbeds that mitigate the costs, rate limits, and reproducibility challenges of live-web interactions.

6 Conclusion

OpenResearcher relocates long-horizon search and browsing to a controllable offline environment, reducing reliance on proprietary infrastructure while retaining explicit information-seeking actions. Its synthesized trajectories support post-training and targeted analysis of search behavior, evidence interaction, and retrieval success.

  • Offline search-and-browse makes long-horizon trajectory synthesis more reproducible, less costly, and less dependent on proprietary infrastructure.
  • The explicit search, open, and find abstraction models realistic information-seeking behavior through a simple browser interface.
  • Trajectories synthesized with OpenResearcher are effective for post-training open-weight deep research agents on fixed-corpus and live-web benchmarks.
  • The controlled setup enables analyses of search behavior, evidence interaction, and the relationship between retrieval success and final answer accuracy.
  • Corpus embeddings use Qwen3-Embedding representations indexed with Faiss, requiring approximately eight hours on 8 A100 80G GPUs and 4 H100 80G GPUs for indexing.
  • GPT-OSS-120B generates trajectories with up to 128K tokens and 150 turns, while synthesis is parallelized across 64 H100 GPUs and takes around 2 days.

A.3 Evaluation Details

The evaluation spans controlled closed-web and live open-web deep-research benchmarks, with tool-augmented proprietary baselines and open-weight research-agent comparisons. It also examines solution coverage as sampling budgets increase.

  • Evaluation Datasets: BrowseComp-Plus uses a fixed corpus with human-verified supporting documents and hard negatives to evaluate deep retrieval and multi-hop reasoning.
  • Evaluation Datasets: BrowseComp, GAIA, and xbench-DeepSearch evaluate persistent browsing, real-world tool use, and sustained multi-turn information seeking on open-web tasks.
  • Compared Baselines: Proprietary baselines are equipped with search tools under the BrowseComp-Plus evaluation setup.
  • Increasing the sampling budget from Pass@1 to Pass@16 consistently improves solution coverage across unique queries.
  • Compared Baselines: Compared open-weight baselines include DeepMiner-32B, CutBill-30B-A3B, and Nemotron-3-Nano-30B-A3B, covering supervised fine-tuning, reinforcement learning, and the student base model.

B.1 More Synthetic Trajectory Analysis

Synthetic trajectory analysis shows that higher sampling budgets improve solution coverage but leave a difficult subset unresolved. The appendix also specifies the answer-processing and browser-tool interfaces used to construct and evaluate these trajectories.

  • More Synthetic Trajectory Analysis: Increasing sampling from Pass@1 to Pass@16 yields consistent performance gains and exposes the model’s broader solution coverage.
  • More Synthetic Trajectory Analysis: A subset of queries reaches a performance plateau because early failures remain unsolved with increased sampling budgets.
  • Data and Evaluation Interfaces: Answer normalization extracts short final answers, including content inside boxed expressions, for evaluation and corpus bootstrapping.
  • Data and Evaluation Interfaces: The response format separates explanation, exact answer, and confidence, while the judge compares extracted answers with reference answers and permits small numerical error margins.
  • Data and Evaluation Interfaces: The browser interface includes search, open, and find operations for retrieving results, opening pages, and locating exact patterns.
  • Case Study: A case study reports that OpenResearcher resolves one query in 5 tool calls by combining search, open, and find.

D.2 Case 2: Deep Research on BrowseComp-Plus

OpenResearcher resolves a convoluted BrowseComp-Plus query by iteratively refining hypotheses and using open and find to verify evidence, ultimately identifying the June trial start.

  • Case setup: The query links a homicide affecting a music group’s parent, an attorney’s prior case, and a young trial witness.The reasoning traces progressively connect these constraints across multiple sources.
  • Trajectory: 14 search + 9 open + 1 find constitute the successful case study’s tool usage.The case demonstrates a long-horizon search-and-verification trajectory.
  • Search refinement: The initial search for a murdered singer’s parent produces results that do not obviously contain the needed information.The model therefore reformulates the query rather than accepting the first results.
  • Evidence verification: The model uses open to inspect candidate pages and find to verify dates, identities, and relationships across sources.This includes confirming the attorney connection, the homicide details, the Jackson 5 relationship, and the young witness.
  • Outcome: June is the exact answer for the month when the trial began.The final answer identifies the case as the murder of Dee Dee Jackson and reports a June 12, 1998 start date.

D.3 Case 3: Deep Research on BrowseComp

In a BrowseComp case, OpenResearcher resolves ambiguous clues about an athlete-turned model through iterative search refinement, page inspection, and cross-validation, identifying Agent Hamilton.

  • Case setup: The case asks for a TV series associated with an athlete-turned model who worked at his father’s restaurant and debuted for a brand founded in the 1950s.The clues constrain the person, modeling debut, brand history, and television credit.
  • Trajectory: 8 search + 11 open + 4 find constitute the successful case study’s tool usage.The trajectory combines search, page navigation, and in-page search over a long reasoning chain.
  • Search refinement: Initial searches for the restaurant clue return unhelpful or ambiguous results, prompting query refinement.The model considers Younes Bendjima as a candidate and checks the result rather than relying on the search snippet.
  • Candidate verification: Page inspection identifies Younes Bendjima as a model, actor, entrepreneur, and former boxer associated with Givenchy and Agent Hamilton.The evidence supports matching the biographical and television clues.
  • Constraint checking: Find verifies that Givenchy was founded in 1952, satisfying the brand-year constraint.The brand’s founding date falls within the required 1950s range.
  • Outcome: Agent Hamilton is the exact answer identified for the requested TV series.The final answer names Younes Bendjima and his notable acting credit.

D.4 Case 4: Deep Research on GAIA

In a GAIA case, OpenResearcher combines targeted search, page navigation, and in-page search to trace an animated song back to its composer, Roger Miller.

  • Case setup: The question asks who composed a song performed by a rooster and a hamster in separate animated videos with different tempos and lyrics.The task requires connecting the two video representations to one underlying song.
  • Trajectory: The trajectory uses 8 search + 11 open + 4 find calls.It demonstrates a complete search–open–find reasoning cycle.
  • Search and identification: The model identifies the hamster video as the Hampster Dance and connects its song to Whistle-Stop from Disney’s Robin Hood.The reasoning narrows from the viral video to the sampled source song.
  • Evidence verification: Find confirms that the Hampsterdance hook uses a sped-up sample of Whistle-Stop.The page also identifies the sample’s origin in the 1973 Disney film Robin Hood.
  • Outcome: Roger Miller is the exact answer for the composer of Whistle-Stop.The final answer states that Miller wrote and originally performed the song.

D.5 Case 5: Failure from Reasoning Error (Gold Document Found, Wrong Conclusion)

The case studies show two distinct failure modes: incorrect reasoning after retrieving the gold document and an empty answer when browser navigation tools are unavailable.

  • Gold document found, wrong conclusion: The Tri-Rail failure retrieves the correct FDOT document but misreads its table columns, selecting the wrong station time.The error concerns reasoning over evidence rather than document retrieval.
  • Failure comparison: The contrast separates retrieval failure from reasoning failure: one case finds the evidence but misinterprets it, while the other cannot navigate beyond search.Together, the cases motivate explicit browser actions and careful evidence reasoning.
  • Missing browser tools: The missing-tools failure provides only search, leaving the model unable to inspect or locate information within pages.The example illustrates why browser navigation tools are needed beyond search results.
  • Missing browser tools: After one irrelevant search result, the model spends 98 subsequent reasoning turns without executing another tool call and produces an empty answer.The trajectory repeatedly formulates new queries but never advances the browser interaction.

D.7 Case 7: Failure from Missing Browser Tools (Cannot Find Gold Document)

A search-only agent fails to identify the gold document on a niche artist query, despite extensive searching and reasoning. Without document-opening and in-page search capabilities, it shifts toward speculative guesses and returns an incorrect answer.

  • 64 search calls and 87,000 reasoning tokens failed to locate the gold document.
  • The agent identified several constraints early but could not navigate to the gallery or biography page containing decisive evidence.
  • Without open and find, the agent increasingly relied on speculative guesses, eventually converging incorrectly on J. Cole.
  • Search snippets were insufficient when the gold document was not prominently indexed, motivating explicit browser navigation tools for long-tail queries.
  • The failure case used 64 search calls, with no open or find operations.
Loading 2603.20278v1…