license: cc-by-4.0
pretty_name: MiniWeb agent trajectories
task_categories:
- other
tags:
- web-agents
- browser-agents
- trajectories
- benchmark
size_categories:
- 10K<n<100K
MiniWeb agent trajectories
Every agent episode behind the results of Back to Basics: A Skill-Centric Approach to Generalist Web Agents: 12,245 episodes in 62 runs, with 375,447 screens. Browse them at minhphd.github.io/miniweb/trajectories.html.
| Benchmark | In the paper | Agents or arms | Runs | Episodes |
|---|---|---|---|---|
| MiniWeb, all 369 tasks | Tab. 2 | 10 | 10 | 3,689 |
| MiniWeb held-out tasks | Tab. 3 | 5 | 15 | 930 |
| MiniWeb single primitives | Fig. 7 | 4 | 12 | 1,236 |
| WebArena-Lite | Tab. 3 | 5 | 15 | 2,310 |
| WebVoyager | Tab. 3 | 5 | 5 | 2,680 |
| Online-Mind2Web | Tab. 3 | 5 | 5 | 1,400 |
Benchmarks: MiniWeb, WebArena-Lite, WebVoyager and Online-Mind2Web. Each episode's success is the one the paper counts: the VLM judge on all 369 MiniWeb tasks (Tab. 2), the task verifier on MiniWeb's held-out tasks and single primitives, WebArena's evaluator, the WebVoyager judge and WebJudge on Online-Mind2Web.
Layout
index.json benchmarks and runs: arm, seed, episodes, success rate, screens
runs/<run>/index.json the run's episodes: task, instruction, outcome, step and screen counts
runs/<run>/<episode>.json one episode
A run id is <benchmark>/<arm>[/s<seed>], for example heldout/miniwebagent/s1. An episode file holds the task and
its instruction, success, graded_by, the grader's verdict, per-primitive verdicts with the judge's reasons
(primitives, where the run was judged per primitive), the planner's subtasks (plan), and steps: each with the
page URL, the action, the action's result or error, the agent's reasoning, and image, an index into images (WebP
data URIs, 960 px wide).
Screens
Screens are the ones the agent saw, where the run kept them. The planner harness on MiniWeb kept none; for those
22,789 steps the screen is redrawn on MiniWeb's sites. On the held-out tasks it is rendered from the page HTML
the recorder saved before the step, with the sites' own stylesheets and images (redrawn: true): close to what the
agent saw, though scripts do not run and pages built on the fly may miss an image. On the single-primitive tasks the
recorder saved no HTML, so the screen is a fresh visit to the step's address (redrawn: "url"): the right page, but
without typed text or other in-page state. WebArena-Lite planner runs kept neither, so their steps have no screen.
Privacy
Text went through the same scrubbing as the MiniWeb release: annotator names appear only as pseudonyms, the hosts
the benchmarks ran on are replaced (http://shopping.webarena, http://map.webarena, ...), and the eval machines'
network addresses and file paths are masked (x.x.x.x, /downloads). Screens of pages that print the visitor's
address (search engines' unusual-traffic pages, error pages with the client IP, an ad page that geolocates the
visitor) are withheld: 452 steps, marked withheld: true. Live websites saw the eval machines' location,
so some pages show it. The MiniWeb sites are synthetic; the live-web runs visited public websites.
Citation
@misc{miniweb2026,
title = {Back to Basics: A Skill-Centric Approach to Generalist Web Agents},
author = {Pham-Dinh, Minh and Ishmam, Md Farhan and Yupanqui, Hernan and Joarder, Reaz Hassan and Marino, Kenneth},
year = {2026}
}