WebResearcher: Unleashing Unbounded Reasoning for Long-Horizon Agents

Abstract

WebResearcher is a novel framework for building long-horizon agents, realized through two key components:

1. IterResearch — an iterative deep research paradigm that reformulates deep research as a Markov Decision Process. The agent periodically consolidates findings into an evolving report while maintaining a focused workspace, overcoming context suffocation and noise pollution in existing single-context methods. 2. WebFrontier — a scalable data synthesis engine that generates high-quality training data via tool-augmented complexity escalation, systematically creating research tasks that bridge the gap between passive knowledge recall and active knowledge construction.

Notably, training data from this paradigm also significantly enhances tool-use capabilities of traditional single-context approaches, and the paradigm extends naturally to parallel thinking for multi-agent concurrent exploration. Extensive experiments across six challenging benchmarks show WebResearcher achieves state-of-the-art performance, even surpassing frontier proprietary systems.

Key points

• Problems with the single-context paradigm used by most deep research agents (WebThinker, WebShaper, WebSailor, etc.):
1. Cognitive workspace suffocation — the expanding context crowds out the model's active reasoning space, forcing premature conclusions. 2. Irreversible noise pollution — without filtering or revision, irrelevant information and early errors persist and compound.

IterResearch

• Reformulates deep research as an MDP with periodic state reconstruction. Each round's state contains only: the research question, the evolving report (central memory synthesizing all prior findings), and the most recent action/tool response.
• Each turn produces a structured Think–Report–Action triple:
• Think: cognitive scratchpad for reasoning and planning (not carried forward).
• Report: the evolving central memory — a coherent, high-density synthesis of new findings with existing knowledge.
• Action: a tool call (e.g., search, code interpreter) or the final answer.
• Maintains a constant-size workspace regardless of research depth, enabling error recovery via report revision and theoretically unbounded research depth.

WebFrontier (data synthesis engine)

Three-stage iterative workflow: 1. Seed data generation from a diverse multi-disciplinary corpus (web pages, papers, e-books), forming composite units for multi-source QA seeds. 2. Iterative complexity escalation — a tool-augmented agent (web search, scholar search, browser, Python) evolves QA pairs via knowledge expansion, conceptual abstraction, fact grounding through cross-validation, and computational formulation. 3. Strict quality control — questions solvable by a tool-free baseline agent are filtered as too easy; questions unsolvable by a tool-augmented advanced agent are discarded or flagged for expert review; a similarity scorer removes semantic redundancy.

Training and test-time scaling

• Rejection-sampling fine-tuning (RFT): keep only trajectories whose final answers exactly match references; train on conditional log-likelihood of structured responses given states, preserving the Markov property.
• Reinforcement learning: each trajectory decomposes into per-round training samples (substantial data amplification vs. one sample per trajectory for single-context methods); optimized with Group Sequence Policy Optimization (GSPO) and minimum-loss downsampling (<1% data loss) for distributed stability.
• Research-Synthesis framework (test-time scaling): n agents solve the problem in parallel via IterResearch; a single synthesis agent (Qwen3-235B-A22B) integrates the final reports and answers into a comprehensive conclusion — feasible because each report concisely encapsulates its full reasoning path.

Experimental setup

• Backbone: Qwen3-30B-A3B.
• Tools: Search (Google), Scholar (Google Scholar), Visit (URL content extraction with goal-directed summarization via Jina + Qwen3), Python sandbox.
• Benchmarks: HLE (Humanity's Last Exam), GAIA, BrowseComp-en, BrowseComp-zh, Xbench-DeepSearch, FRAMES.
• Baselines: general tool-augmented LLMs (Qwen3, Claude-4-Sonnet, OpenAI-o3, DeepSeek-V3.1/R1, GLM-4.5, Kimi-K2), commercial deep research agents (OpenAI Deep Research, Gemini Deep Research, Perplexity, Grok, Kimi-Researcher), and open-source agents (WebDancer, WebSailor, MiroThinker, WebExplorer).
• Metric: pass@1 with LLM-as-a-Judge; nucleus sampling with temperature 0.6, top-p 0.95.

Main results

| Backbone | HLE | BrowseComp | BrowseComp-ZH | |---|---|---|---| | Qwen3-30B-A3B | 13.2 | 0.5 | 13.5 | | Qwen3-235B-A22B | 20.0 | 2.3 | 29.4 | | DeepSeek-R1 | 24.8 | 8.9 | 35.7 | | Claude-4-Sonnet | 20.3 | 12.2 | 29.1 | | OpenAI Deep Research | 26.6 | 51.5 | - | | DeepSeek-V3.1 | 29.8 | 30.0 | 49.2 | | WebResearcher-30B-A3B | 28.8 | 37.3 | 45.2 | | WebResearcher-30B-A3B-heavy | 36.7 | 51.7 | 56.8 |

• On HLE, WebResearcher-heavy achieves 36.7% — beating DeepSeek-V3.1 (29.8%), OpenAI Deep Research (26.6%), and Gemini Deep Research (26.9%), a 6.9-point improvement validating sustained deep reasoning across extended research.
• On BrowseComp-en, WebResearcher-heavy reaches 51.7%, on par with OpenAI Deep Research (51.5%) and 21.7 points above the best open-source alternative (DeepSeek-V3.1, 30.0%).
• On BrowseComp-zh, WebResearcher-heavy hits 56.8%, close to o3 (58.1%) and well above DeepSeek-V3.1 (49.2%).

Conclusion

WebResearcher demonstrates that iterative synthesis with periodic state reconstruction fundamentally outperforms linear context accumulation for long-horizon deep research, delivering state-of-the-art results with an efficient 30B-A3B backbone and closing the gap to — and exceeding — frontier proprietary systems.