WebResearcher: Unleashing Unbounded Reasoning for Long-Horizon Agents

Key points

• Problem with single-context deep research: Existing open-source systems (WebThinker, WebShaper, WebSailor) accumulate all retrieved information and reasoning into one expanding context window. This causes two failures:
1. Cognitive workspace suffocation — the fixed context budget is dominated by historical data, crowding out deep reasoning and forcing premature conclusions. 2. Irreversible noise pollution — irrelevant information and early errors persist without filtering or revision, compounding bias over time.
• IterResearch: Reformulates deep research as a Markov Decision Process with periodic state reconstruction. Each round's state contains only the research question, an evolving report (the agent's central memory), and the most recent tool interaction. Each step produces a Think–Report–Action triple. The report is actively synthesized — new findings are integrated with existing knowledge, conflicts resolved, noise filtered — keeping the workspace constant-sized regardless of research depth, enabling error recovery via report revision and theoretically unbounded research depth.
• WebFrontier (data engine): A scalable synthesis pipeline with three stages:
1. Seed generation — QA pairs created from composite units of multi-disciplinary documents (web pages, papers, e-books), requiring multi-source synthesis. 2. Iterative complexity escalation — a tool-augmented agent (web search, academic 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 even by a tool-augmented agent are discarded or flagged for human review; a similarity scorer removes redundancy.
• Training:
• Rejection-sampling fine-tuning (RFT): keeps only trajectories whose final answer exactly matches the reference; the objective maximizes conditional log-likelihood of each round's response given the previous state, enforcing the Markov property.
• Reinforcement learning with GSPO: each trajectory naturally decomposes into per-round training samples (unlike single-context methods that yield one sample per trajectory), giving large data amplification. Minimum-loss downsampling keeps distributed batches fixed-sized with <1% data loss.
• Research-Synthesis (test-time scaling): n parallel IterResearch agents solve the problem independently; a single synthesis model (Qwen3-235B-A22B in experiments) integrates their final reports. Because each report compactly encapsulates a full reasoning path, the synthesizer can consider many diverse solution strategies within a bounded context.
• Tools: Search (Google, batched queries, top-10 results), Scholar (Google Scholar metadata), Visit (Jina-based page retrieval plus goal-directed summarization via Qwen3), and Python (sandboxed execution).

Main results (backbone: Qwen3-30B-A3B)

| Model | HLE | BrowseComp-en | 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-heavy | 36.7 | 51.7 | 56.8 |

• On HLE, WebResearcher-heavy reaches 36.7%, beating DeepSeek-V3.1 (29.8%), Gemini Deep Research (26.9%), and OpenAI Deep Research (26.6%).
• On BrowseComp-en, WebResearcher-heavy (51.7%) matches OpenAI Deep Research (51.5%) and exceeds the best open-source baseline (DeepSeek-V3.1, 30.0%) by 21.7 points.
• On BrowseComp-zh, it reaches 56.8%, close to o3 (58.1%) and well above DeepSeek-V3.1 (49.2%).
• Benchmarks used: HLE, GAIA, BrowseComp-en/zh, Xbench-DeepSearch, FRAMES; evaluation via pass@1 with LLM-as-a-Judge, temperature 0.6, top-p 0.95.

Notable findings

• Training data generated under the IterResearch paradigm also improves tool-use capability of traditional single-context methods.
• Iterative synthesis preserves reasoning quality across 10 or 100+ rounds, where single-context systems suffer diminishing returns from context bloat.
• Parallel thinking (multi-agent concurrent exploration plus ensemble synthesis) yields more comprehensive conclusions than a single trajectory.