hexuan21/gui-env-shaping-m2w-all300-actor2

GUI env shaping: Online-Mind2Web, all 300 tasks, actor v2

Trajectories of a web agent on all 300 tasks of Online-Mind2Web, run on 2026-10-22 on a US server, headless Chromium, 12 parallel workers. Companion to gui-env-shaping-m2w-test120, which used the previous actor.

Methods

method	What the actor sees
no_shaping	The raw screenshot
shaper_no_hist	Screenshot modified each step by an LLM shaper (som / highlight / mask / zoom / grid / none) plus a text hint
shaper_hist	Same, and the shaper also sees its last two steps (image, its own intervention and hint, the actor's action and whether the URL changed)

The actor, the shaper and the judge (WebJudge) are all Qwen3.6-27B. Each episode has at most 15 steps.

Actor v2

Compared with the actor in the test120 dataset, the actor

1. sees what each past action did: execution status, whether the URL changed, whether the page visibly changed, and after type the text now in the focused field (result in steps.jsonl);
2. replaces the text of the focused field when it types ("append": true to add); typing with no field in focus is not executed and reported as FAILED;
3. answers with check (did the previous action do what it expected), progress, suggestion (follow or ignore the hint, and why), thought, expect and action;
4. gets the shaper's hint as "a suggestion that may be wrong", with rules to close pop-ups first and to trust the screenshot over the hint;
5. also gets the unmodified screenshot when the view is mask or highlight (raw_image: true).

Coverage

no_shaping has 300 episodes, shaper_no_hist 299 and shaper_hist 296. The missing ones are theweathernetwork.com tasks: from 16:55 that day the site answered the headless browser with a redirect loop (ERR_TOO_MANY_REDIRECTS) and the episodes could not start.

Layout

index.jsonl                                one row per (method, task)
episodes/<method>/<task_id>/meta.json      task, website, level, score, final answer, judge reasoning, error cause
episodes/<method>/<task_id>/steps.jsonl    per step: url, iv, thought, action, status, check, progress,
                                           suggestion, expect, result, raw_image
episodes/<method>/<task_id>/step_XX.jpg    annotated screenshot of step XX

Image annotations

• The image is what the actor saw: the shaped view for shaper runs, the raw screenshot for no_shaping. For mask / highlight steps the actor also saw the raw screenshot, which is not repeated here.
• Orange crosshair: where the executed click landed.
• Bottom panel: step number, intervention type and chosen elements, the shaper's hint, the action and its execution status, and the result line the actor saw at the next step.
• som draws boxes only on the elements the shaper chose and renumbers them #1..#k on the image. The el=[...] in the panel are the ids in the page's full element list.

Fields

• score: WebJudge label, 1 = success.
• error_cause (failed episodes only): primary failure cause assigned automatically by an LLM from the trajectory and the judge's reasoning. Not checked by hand. Values: shaper_loop, shaper_misleading_hint, shaper_render, actor_planning, actor_wrong_answer, actor_grounding, env_blocked, judge_or_task.
• no_shaping_solved: whether the no_shaping run of this dataset solved the same task.