hexuan21/gui-env-shaping-m2w-test120

GUI env shaping: Online-Mind2Web test split (120 tasks)

Trajectories of a web agent on the 120 test tasks of a level-stratified 6/4 split of Online-Mind2Web (300 tasks, seed 0).

Methods

method	What the actor sees
no_shaping	The raw screenshot
shaper_no_memory	Screenshot modified each step by an LLM shaper (no learning-state estimate)
shaper_memory	Same, and the shaper also reads a store of past trajectories to estimate a learning level

The actor is Qwen3.6-27B (VLM) in all three. The shaper is Qwen3.6-27B as well. Each episode has at most 15 steps and is scored by WebJudge (Qwen3.6-27B).

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, intervention (iv), thought, action, status
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.
• 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 (SUCCESS / FAILED).
• 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 (shaper runs, 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 solved the same task.