multimodal-ai-engineering/preflight-bench

Preflight Bench

Screenshots and labels for testing the check a computer-use agent runs right before it clicks. Data for the paper What Does "Now" Mean to Your Computer-Use Agent? Stale Screenshots, Page Loads, and Which Check to Run Before a Click (Srinivasan, 2026). Code, generator and preregistrations: github.com/multimodal-ai-engineering/preflight-bench.

An agent plans on a screenshot and clicks a few seconds later. Each example here pairs the screenshots the agent planned on with the screenshots at the moment of the click, and records from the page itself whether the planned click would still land on its target. The pages are six synthetic web pages that move by design (a form, a market dashboard, a news feed, a shared document, a map and a drawing board); all content on them is invented.

Tables

The dataset is stored in Lance format, so images, labels and model answers sit in the same tables and can be filtered without downloading everything.

Table	Rows	One row is
click_pairs	4,608	a planned click: two screenshots at planning time (200 ms apart), two screenshots 6.5 s later, the planned box and click point, and the ground truth
outcome_pairs	3,600	an action and its result: two screenshots before, two after, the expected effect in words, and whether the action failed
checker_answers	31,610	one checker's answer to one test pair: pixel checks, PUSV, Qwen3-VL 2B and 4B, SmolVLM2, Florence-2, Claude Haiku 5.5 and Claude Opus 5.5

Key columns in click_pairs: stale (the planned click would no longer land on its target), dialog (a dialog appeared), event (what the generator changed: none, moved, moved_lookalike, covered, gone, dialog, slow_load), page, tier (0 static form, 1 moving pages, 2 remote-desktop conditions), split, truth_json (the page's state at nine moments after the click), and the four screenshots as PNG bytes (planned_a, planned_b, current_a, current_b).

Key columns in outcome_pairs: failed (the expected effect did not appear), outcome (success_clear, success_small, failed_error or failed_silent), expected_effect, effect_box, and the screenshots before, before_b, after, after_b.

checker_answers is in long format: pair_id, question (action for the click questions, worked for the "did it work?" question, and the sub-questions a small model was asked when the pixel check escalated), checker, answer (the checker's raw output) and latency_ms where it was measured per pair.

Splits

Layouts 0 and 1 of each page are dev (used for tuning); layouts 2 and 3 are test (held out; all thresholds in the paper were frozen before test labels were read). Labels for both splits are included.

Load it

import io, lancedb
from PIL import Image

db = lancedb.connect("hf://datasets/multimodal-ai-engineering/preflight-bench/data")
pairs = db.open_table("click_pairs")

stale = (pairs.search()
         .where("split = 'test' AND page = 'live_newsfeed' AND stale")
         .select(["pair_id", "event", "planned_box", "planned_a", "current_b"])
         .limit(5).to_pandas())
Image.open(io.BytesIO(stale.planned_a[0])).show()

To work offline, download the repository and connect to the local data folder instead.

What is not included

Screenshots of the 31 public web pages used in the paper's real-page measurements are not released, because they contain third-party content. The real-page results are in the GitHub repository as statistics.

Citation

@article{srinivasan2026preflight,
  author = {Srinivasan, Vasundra},
  title  = {What Does ``Now'' Mean to Your Computer-Use Agent? Stale Screenshots, Page Loads, and Which Check to Run Before a Click},
  year   = {2026},
  url    = {https://github.com/multimodal-ai-engineering/preflight-bench}
}

License: CC BY 4.0. Independent research, not affiliated with or endorsed by any employer.