| --- |
| license: cc-by-4.0 |
| tags: [human-study, web-agents, grounding, hci] |
| --- |
| |
| # PageGuide user study — results |
|
|
| One row per answered task from the PageGuide web study, on the task selection the analysis |
| dashboard uses by default. `rows.csv` is the row-level data; the two `behavior_*.csv` files are |
| the aggregates plotted in the behavioural figures (mean, sd, se and n per cell). `outcomes.csv` |
| carries accuracy, localization F1 and its two halves, and the timing stages — mean, sd, se, |
| median and quartiles per cell. |
|
|
| Exported 2026-08-14T05:06:12.478Z · 217 rows · 43 participants. |
|
|
| ## Conditions |
|
|
| Every participant sees both arms, interleaved task by task: |
|
|
| - `nongrounding` — the agent reports an answer with no evidence attached. |
| - `grounding` — the same claims, with citations into the page and saved image crops. |
|
|
| ## Task selection |
|
|
| Some tasks are excluded from their card by default. The figures and these files use the same |
| selection, so they agree with the dashboard: |
|
|
| | Facet | Left out | Why | |
| |---|---|---| |
| | Find × Text | MARS-v1 | a disputed answer key | |
| | Find × Visual | MUFC-V1, PEDANT-V1 | two runs whose grounded locate time is ~14× the others while accuracy and localization stay flat | |
| | Guide × Text | gv2-ed05a7b6-kk24zp, gv2-ms9iw0pq-5kj5zr, gv2-msf02a2n-88li4p, gv2-msf5mo9m-qm5brt | three duplicate re-recordings, plus both recordings of the New York goal | |
| | Guide × Visual | gv2-msf1pyqv-omt0hz | a run whose ground truth blames one step under two error types, so no participant can score it in full | |
|
|
| ## What is not here |
|
|
| Participants' free-text notes are omitted, and session ids are replaced with a per-export |
| `participant` integer. Both are deliberate: the notes are the only column that can carry |
| something about a person rather than about a task. |
|
|
| ## Scoring |
|
|
| Localization quality is F1. For Find it is over the passages picked; for Guide it is the mean of |
| an F1 over the error types named and an F1 over the steps blamed, with a correct "no error" |
| scoring in full on a run that contains none. A null precision means the participant predicted |
| nothing, which is zero true positives, so its F1 is 0 rather than missing. |
|
|
