WebResearcher condenses each round into a report instead of growing its context

Most open work on deep search runs on one simple principle: pile everything you find into a single large context window. Each step adds new excerpts, links and notes. Eventually the useful material drowns in noise, early mistakes stay in the record forever, and the room left for thinking shrinks fast. The authors of WebResearcher propose the opposite: periodically stop the stream, squeeze what matters into a compact report, and start the next round with a clean desk. The agent keeps the knowledge without hauling the baggage forward.

The core idea: research in rounds

WebResearcher is built around IterResearch, an iterative scheme framed as an MDP. Each round has three parts:

• Think: a brief deliberation, the plan for the step. It lives only inside the current round.
• Report: the central memory — a clean distillation of everything already checked and reconciled.
• Action: a tool call or the final answer.

Only the updated Report and the last tool response carry over between rounds. All the ephemeral mess is discarded. The context never swells, and the reasoning stays clear and stable even across dozens of steps.

Where hard tasks come from: the WebFrontier data factory

To teach an agent to actually do research, you need tasks that demand more than recalling a fact — tasks whose answers have to be built step by step. So the authors created WebFrontier, a scalable pipeline for generating and filtering them.

• First, dense fragments are collected from the web, papers and books, and combinations of those fragments seed the initial questions and answers.
• Then the same tool-equipped agent (web and Google Scholar search, browser extraction, Python) makes the tasks progressively harder, adding calculations and cross-checks.
• Finally comes strict quality control: easy pairs are thrown out, hard ones are solved by a strong solver, duplicates are removed, and facts are verified by judges. What remains is a body of verifiable, genuinely difficult tasks — exactly the kind where the gap between a bare LLM and a system with tools shows up sharply.

How it is trained and how it scales at inference

The model is trained on correct IterResearch trajectories: only those whose result matched the reference are kept, and the model learns the Think–Report–Action template. This helps separate reasoning from noisy tool output. RL comes next: trajectories are cut into rounds, which yields far more training examples than a mono-context scheme does.

At deployment the system uses a technique called Research-Synthesis. Several independent agents explore the task in parallel, each producing a final report and an answer. A separate synthesizing agent then aggregates the reports rather than the full histories. That packs more distinct lines of reasoning into a limited context window, and the result comes out steadier.

How it does on benchmarks

The system is built on Qwen3-30B-A3B with a toolset of Search, Scholar, Visit and Python. Across six benchmarks, WebResearcher matches the best systems or beats them:

• HLE: 36.7%, ahead of DeepSeek-V3.1 (29.8%) and OpenAI Deep Research (26.6%).
• BrowseComp-en: 51.7% — level with OpenAI Deep Research (51.5%).
• BrowseComp-zh: 56.8%, against 58.1% for o3 and 49.2% for DeepSeek-V3.1.
• GAIA: 72.8% — above Claude-4-Sonnet (68.3%) and o3 (70.5%).
• Xbench-DeepSearch: 71.0% — close to DeepSeek-V3.1 (71.2%).
• FRAMES: 84.8% — above DeepSeek-V3.1 (83.7%) and o3 (84.0%).

The ablations show the main gain comes from iterative inference specifically: mono-context degrades as chains get longer, while IterResearch holds up steadily over long routes. Tool profiles adapt to the task: on HLE, Scholar dominates and chains stay short; on BrowseComp it is long navigation with Search and Visit, averaging more than 60 steps.

How far parallelism goes

The authors checked how quality scales with the number of parallel researchers n. Up to n=8 the gain is clear; past that, returns fall off while cost keeps rising linearly. The practical balance sits around eight independent trajectories.

Why this matters

The main contribution is a change of paradigm. Instead of piling up context without limit, the agent consolidates what it knows into a compact report and rebuilds its workspace from scratch. That arrangement fits how a person runs a long investigation: drafts, a summary, the next pass. On top of that, the WebFrontier data factory teaches the model to construct an answer step by step rather than recall it, which is what tasks involving several sources, calculations and checks require. The result is more reliable reasoning, better tool use, and stability over very long chains.

💾 Code