A 7B agent that drives a real browser beats Search-R1 by 20%

Most of today's web agents solve tasks through a long pipeline: scrape the page, compress it into text, hand it to an LLM. That is convenient, but poor in actions: no real scrolling, no clicks, no work with tabs or forms. Costs climb too, because of all the external calls. The BrowserAgent team proposes going back to the original source — acting directly in the browser, the way a person does. That opens the door to deeper exploration of pages and better multi-step reasoning.

What the authors actually propose

BrowserAgent is an agent that sees a live page and drives it through Playwright with a small set of atomic actions: click, scroll, type, navigate to a URL, manage tabs, and stop with an answer. No separate parsers or summarizers — the model learns to “read” the page representation available to it and make decisions on the fly. Inside runs a think–summarize–act loop: at every step the agent draws conclusions and writes them into an explicit memory, so that important facts are not lost a few screens later.

How it works under the hood

The main engineering obstacle is the speed of the browser environment. It normally yields 1–2 episodes per minute, which makes data collection expensive. The authors build their orchestration on Ray and run dozens of Playwright instances in parallel on a single 32‑core server. The result is 50+ episodes per minute and a more than tenfold drop in the cost of collection. Sessions are logged in detail: the prompt, the page observation, the reasoning, the action, and the key intermediate conclusions committed to memory. The infrastructure sits behind a single interface built on FastAPI. Wikipedia is served locally through Kiwix for stability and reproducibility.

Where the training data comes from

The authors collect 5.3K high-quality trajectories on basic and multi-hop questions: NQ and HotpotQA. Simple cases are capped at 6 steps, hard ones at 30. The data is rich in context: every step carries the observation, the reasoning, and whatever went into memory. That format teaches the model to move carefully across pages and to stitch facts together over several hops.

Training without heavy RL

The recipe is simple and practical: supervised fine-tuning first, then rejection fine-tuning. The base model is Qwen2.5‑7B‑Instruct. In the SFT stage the model is taught the answer format and a basic strategy. In the RFT stage the authors keep the trajectories where several samples contain both wrong and right answers, and from the right ones take those with deeper reasoning. A share of the original SFT examples is mixed back in so the model does not lose its discipline about format. The result is stronger reasoning without complicated RL and without enormous datasets.

What the experiments show

Testing runs on six datasets: NQ, PopQA, HotpotQA, 2WikiMultiHopQA, Musique, Bamboogle. The metrics are the familiar EM plus an LLM-based score: an answer counts as valid if at least two of three judge models (GPT‑4.1, Gemini Flash 2.5, Claude Sonnet 3.7) agree that it is correct. That matters, because wordings often diverge from the reference while the meaning is right.

Key findings

• BrowserAgent‑7B delivers roughly a 20% gain over Search‑R1 while training on less data. Direct interaction with the web and the memory turn out to be the decisive parts.
• More steps available means higher accuracy: mean EM rises from ~0.34 with a 6-step limit to ~0.41 with a 30-step limit. The agent simply needs the time to walk the chain of pages.
• On datasets like TriviaQA the gap between EM and the LLM score is plain: the model often gives the right answer in different words. The cases in the appendix show the meaning matching even when the string does not.
• Environment throughput is a win of its own: 50+ episodes per minute against the usual 1–2.

A few illustrative examples

• The question about the “princes in the Tower”: the agent opens the relevant page, picks out the father and answers King Edward IV. Right in substance, even though the reference string is Edward IV of England.
• A question about the bands Skin Yard and Ostava: the first is American, the second Bulgarian. The agent checks both pages in turn and returns an answer equivalent to “no”.

Why it matters

BrowserAgent shows that an LLM can pick up a user's behavior in a browser without cumbersome RL and expensive pipelines. Human-like atomic actions plus an explicit memory give it flexibility on real tasks: forms, logins, infinite feeds, moving between tabs. It is a step toward agents that not only read but act.

The environment is still heavier than plain text, and EM suffers from the variety of phrasings. Better verification metrics and richer sources beyond Wikipedia are needed, and open weights and reproducible environments would help the community. But the direction looks promising: a simple training recipe, scalable infrastructure, and a clear gain on hard tasks.