thegovind/blink-cua

blink CUA

Videos, marked screenshots, per-step probabilities, and traces of blink driving ten apps.

For each step, the harness marks clickable boxes on a screenshot and asks blink three typed questions in one request: which box, done yet, and is the next click risky? Web apps supply the boxes through one DOM query in the browser; canvas apps report their own hit regions. No detector, OCR or second vision model is involved. blink sees the marked pixels, task, and past actions, not page code or the right answer. It stops when "done" reaches 60% and asks before a click when "risky" reaches 50%. Text and dropdown values are picked from the task; blink does not write free text. The app's own checker counts a task only if the goal is met with no wrong side effects.

Zero-shot, 690 runs, seeds 1-10, no pop-ups:

Model	Tasks done	Right next click	Risky clicks paused on
blink-27b	199/230 (87%)	84%	100/100
blink-4b	152/230 (66%)	65%	66/76
blink-mimo-9b	133/230 (58%)	69%	72/82

These are original apps built for this showcase, not a public benchmark.

Files

• runs/cua-zero-shot-20260929/<app>/<task>-<seed>-<model>/: video.mp4 (the run at real speed), preview.mp4 (smaller), poster.webp, steps/NNN.webp (the marked screenshot blink saw at each step) and steps.<hash>.json (each step's boxes, probabilities, pick and time, plus the request's own text: earlier actions, screenshot note and questions). Only the runs the Space shows have media.
• pipeline/calendar-0-1-001-raw.webp: the raw, unmarked screen of one recorded step (Team room calendar, task 0, seed 1, step 1), re-rendered from the same app, seed and task. The page listed the same boxes as the trace, and marking it reproduces the screenshot all three models were sent, pixel for pixel. The Space's How one step works shows it.
• gallery.json: the index the Space reads.
• results/results.jsonl: the outcome of all 690 runs; results/config.json: harness version, thresholds and seeds.
• MANIFEST.json: size and SHA-256 of every file.

Seeds 1-3 are replays with one run per server at a time, so their step times are clean; every replay matched the original outcome. Seeds 4-10 ran 16 at a time, so their step times include waiting.

Reproduce

The ten apps and harness are in examples/cua. Follow its README to run a scenario.

Media and traces: CC BY 4.0. Code: Apache-2.0. Weights: non-commercial research and evaluation only; see each model card's license.