Source-linked AI summary
WebResearcher: Unleashing unbounded reasoning capability in Long-Horizon Agents
Zile Qiao, Guoxin Chen, Xuanzhong Chen, Donglei Yu, Wenbiao Yin, Xinyu Wang, Zhen Zhang, Baixuan Li, Huifeng Yin, Kuan Li, Rui Min, Minpeng Liao, Yong Jiang, Pengjun Xie, Fei Huang, Jingren Zhou
TL;DR
Deep-research agents need to construct and synthesize knowledge from external sources, but mono-contextual approaches face context growth and noise. WebResearcher addresses this with iterative report-based research, tool-augmented data synthesis, and parallel exploration, achieving state-of-the-art results across 6 benchmarks while showing a compute–accuracy trade-off.
Problem
Models can recall passive knowledge but struggle to actively discover, verify, and synthesize new knowledge from external sources.
Method
WebResearcher combines IterResearch’s MDP-based periodic report consolidation, WebFrontier’s tool-augmented complexity escalation, and parallel multi-agent research with integrative synthesis.
Results
WebResearcher achieves state-of-the-art performance across 6 challenging benchmarks, including 36.7% accuracy on HLE and 51.7% on BrowseComp-en.
Takeaways & Limitations
The framework supports sustained long-horizon reasoning through iterative synthesis and enables broader conclusions through parallel exploration.
Takeaways & Limitations
Parallel exploration increases computational cost linearly, while performance gains show diminishing marginal returns for n > 8.
Abstract
from arXiv · showhide
Recent advances in deep-research systems have demonstrated the potential for AI agents to autonomously discover and synthesize knowledge from external sources. In this paper, we introduce WebResearcher, a novel framework for building such agents through two key components: (1) WebResearcher, an iterative deep-research paradigm that reformulates deep research as a Markov Decision Process, where agents periodically consolidate findings into evolving reports while maintaining focused workspaces, overcoming the context suffocation and noise contamination that plague existing mono-contextual approaches; and (2) WebFrontier, a scalable data synthesis engine that generates high-quality training data through tool-augmented complexity escalation, enabling systematic creation of research tasks that bridge the gap between passive knowledge recall and active knowledge construction. Notably, we find that the training data from our paradigm significantly enhances tool-use capabilities even for traditional mono-contextual methods. Furthermore, our paradigm naturally scales through parallel thinking, enabling concurrent multi-agent exploration for more comprehensive conclusions. Extensive experiments across 6 challenging benchmarks demonstrate that WebResearcher achieves state-of-the-art performance, even surpassing frontier proprietary systems.
1 Introduction
The paper targets the gap between passive knowledge recall and active discovery, verification, and synthesis from external sources. It introduces WebResearcher through iterative research, scalable data synthesis, and parallel multi-agent exploration, achieving strong benchmark performance.
- Passive knowledge-centric models struggle to actively discover, verify, and synthesize knowledge from external sources.
- IterResearch reformulates deep research as a Markov Decision Process with periodic report consolidation and workspace reconstruction.
- WebFrontier generates high-quality training data through tool-augmented complexity escalation across increasingly complex research tasks.
- Parallel Research lets multiple agents explore concurrently before an Integrative Synthesis agent combines their final reports and answers.
- 6 challenging benchmarks show state-of-the-art WebResearcher performance, including 36.7% accuracy on HLE and 51.7% on BrowseComp-en.
2 IterResearch: An Iterative Deep-Research Paradigm
IterResearch replaces ever-expanding single-context accumulation with discrete research rounds that repeatedly synthesize findings into an evolving report. Each round reconstructs a lean workspace from essential state, preserving continuity while limiting context bloat and noise.
- Mono-contextual research accumulates information in one expanding context, causing cognitive workspace suffocation and noise contamination.
- IterResearch models research as discrete MDP rounds containing the question, evolving report, and recent tool interaction.
- Each round follows a Think-Report-Action structure, with reasoning, synthesis, and a selected next action.
- The next workspace is rebuilt from the question, previous report, and latest tool response rather than retaining ephemeral reasoning and raw accumulation.
- A constant-size workspace and periodic synthesis preserve reasoning capacity across long-horizon investigations while filtering noise.
3 A Scalable Data Engine for Advancing Agentic Intelligence
The scalable data engine constructs complex, verifiable research tasks through seed generation, iterative tool-augmented refinement, and quality control. Its target is the capability gap between baseline models and tool-augmented counterparts.
- The engine uses a three-stage multi-agent workflow: seed data generation, iterative complexity escalation, and rigorous quality control.
- A curated corpus of webpages, academic papers, and e-books is paraphrased, cleaned, and distilled into information-dense chunks.
- The ItemWriter Agent combines related chunks to generate seed QA pairs requiring multi-source information synthesis.
- Tool-augmented refinement expands scope, abstracts concepts, cross-validates facts, and formulates computational or simulation problems.
- A self-bootstrapping loop systematically escalates task complexity by making each refined QA pair the next iteration’s seed.
- Quality control filters baseline-solvable tasks, retains tasks solved by advanced tool-augmented agents, and discards or flags failures.
- The engine aims to generate large volumes of complex, factually correct tasks within the capability gap between baseline and tool-augmented models.
4 Training and Test-Time Optimization
IterResearch is trained on accepted multi-round trajectories and optimized by treating each research round as a training sample. At test time, parallel research agents explore independently, then a synthesis agent consolidates their reports and answers.
- 4.1 Rejection Sampling Fine-Tuning: Rejection sampling fine-tuning trains IterResearch on well-formed trajectories generated in its structured iterative format.The approach uses trajectories produced by prompting large language models to follow the iterative paradigm.
- 4.1 Rejection Sampling Fine-Tuning: Only trajectories whose final answers exactly match the reference are retained, pairing correct reasoning processes with accurate conclusions.
- 4.1 Rejection Sampling Fine-Tuning: Each round depends only on the immediately previous state, while gradients cover model-generated responses rather than tool observations.This enforces the paradigm’s Markov property and separates reasoning from external tool outputs.
- 4.2 Reinforcement Learning: IterResearch trajectories decompose into multiple training samples, increasing data utilization compared with mono-contextual approaches.The decomposition treats each round as a training tuple and supports group-level advantage normalization across variable research depths.
- 4.3 Research-Synthesis: Harnessing Test-time Scaling with IterResearch: The Research-Synthesis Framework runs n independent research agents in parallel and uses one synthesis agent to integrate their reports and answers.Parallel Research explores distinct tools and reasoning paths, while Integrative Synthesis produces a unified final conclusion from the collected outputs.
5 Experiments
WebResearcher is evaluated across diverse benchmarks, datasets, baselines, and tool-supported research settings. It achieves consistently strong results across general web reasoning and complex goal-oriented tasks, outperforming or matching leading systems on several benchmarks.
- Experimental Setup: WebResearcher is implemented with Qwen3-30B-A3B and evaluated across eight challenging benchmarks using iterative research and WebFrontier-generated training data.The reported main results cover six benchmarks, while the experimental setup describes eight benchmarks overall.
- Benchmarks: The evaluation includes HLE, GAIA, BrowseComp-en, BrowseComp-zh, Xbench-DeepSearch, and FRAMES, covering academic competence, tool use, web navigation, and multi-hop reasoning.HLE uses 2,154 text-only questions; GAIA uses 103 text-only validation cases.
- Baselines: The baselines comprise general tool-using language models, commercial deep-research agents, and open-source web-search agents.Commercial systems were not tested across all benchmarks because some were not fully accessible via API.
- Overall Performance: WebResearcher achieves state-of-the-art performance across six challenging benchmarks, surpassing larger models and existing open-source and proprietary deep-research systems.The benchmarks span complex reasoning, web navigation, and long-horizon information-seeking tasks.
- General Web Navigation and Reasoning Benchmarks: 36.7% accuracy on HLE exceeds DeepSeek-V3.1 (29.8%), OpenAI Deep Research (26.6%), and Gemini Deep Research (26.9%).The reported improvement over the next-best system is 6.9 percentage points.
- Complex Goal-Oriented Web Tasks: 72.8% accuracy on GAIA surpasses Claude-4-Sonnet (68.3%) and OpenAI-o3 (70.5%), while FRAMES reaches 84.8% and Xbench-DeepSearch reaches 71.0%.On Xbench-DeepSearch, WebResearcher matches DeepSeek-V3.1 (71.2%) and exceeds WebSailor-72B (55.0%) and Kimi-K2 (50.0%).
6 Analysis
The analysis isolates the iterative paradigm’s contribution, examines adaptive tool use, and evaluates parallel trajectory scaling. Iterative synthesis improves long-horizon reasoning, while increasing parallel trajectories improves performance with rising cost and diminishing returns beyond n > 8.
- 6.1 The Primacy of the Iterative Paradigm: Mono-Agent + Iter outperforms Mono-Agent across HLE, BC-EN, and BC-ZH, showing the benefit of specialized iterative-research training data.The reported scores are HLE 25.4 vs. 18.7, BC-EN 30.1 vs. 25.4, and BC-ZH 40.4 vs. 34.6.
- 6.1 The Primacy of the Iterative Paradigm: WebResearcher scores 28.8 versus 25.4 for Mono-Agent + Iter on HLE, isolating the iterative paradigm’s additional contribution.The comparison uses the same underlying model architecture while contrasting iterative and linear inference strategies.
- 6.1 The Primacy of the Iterative Paradigm: Periodic synthesis and workspace resets mitigate contextual degradation and irreversible error propagation by maintaining focused reasoning contexts.The analysis attributes linear-strategy degradation to excess low-value historical data and persistent early errors.
- 6.2 Analysis of Tool-Use Behavior: On HLE, Scholar accounts for 25.4% of tool calls and tasks average 4.7 turns, indicating concise information retrieval for knowledge-intensive questions.The tool-use analysis focuses on Search, Scholar, Visit, and Python invocation sequences.
- 6.2 Analysis of Tool-Use Behavior: On BrowseComp, Search and Visit comprise over 96% of tool calls, while tasks average 61.4 turns and the hardest exceed 200 interaction turns.The contrast with HLE reflects longer web navigation and multi-page information integration.
- 6.3 Analysis on Reasoning Trajectories in Reason-Synthesis Framework: Increasing parallel trajectories n consistently improves pass@1, with the largest gains occurring when scaling from 1 to 8.Each Research Agent explores a distinct reasoning path, and the Synthesis Agent fuses their final outcomes.
- 6.3 Analysis on Reasoning Trajectories in Reason-Synthesis Framework: Scaling trajectories increases computational cost linearly, while performance gains show diminishing marginal returns for n > 8.The analysis identifies an accuracy–computational-budget trade-off and recommends n = 8 as a compelling balance.
7 Related Work
Related work describes rapid progress in proprietary and open-source autonomous deep-research agents. Proprietary systems establish strong benchmarks but remain opaque, motivating continued open-source development.
- Deep Research: Proprietary deep-research systems have established benchmarks for autonomous research capabilities but remain opaque.The related-work discussion contrasts these systems with open-source efforts.
- Deep Research: Open-source efforts have developed in response to the progress of proprietary deep-research systems.The passage names multiple proprietary and open-source research efforts.
8 Conclusion
The conclusion presents WebResearcher as a framework combining iterative research, scalable data synthesis, and parallel multi-agent exploration. Experiments across six challenging benchmarks report state-of-the-art performance, including results surpassing frontier proprietary systems.
- 8 Conclusion: WebResearcher combines IterResearch, WebFrontier, and a Research-Synthesis Framework as its three key contributions.IterResearch uses periodic consolidation, WebFrontier uses tool-augmented complexity escalation, and Research-Synthesis enables parallel multi-agent exploration.
- 8 Conclusion: Experiments across 6 challenging benchmarks show state-of-the-art performance that surpasses even frontier proprietary systems.The conclusion frames these results as supporting structured iteration with periodic synthesis rather than unbounded accumulation.