# Reading Right, Acting Wrong: Diagnosing Decision Failures of System One Models in Agent Loops

Pairs factual probes with next-action choices on frozen agent traces, including Mind2Web, to diagnose reading-decision gaps without closed-loop execution.

- Catalog URL: https://awesomecua.github.io/resources/arxiv-2610-18579/
- Original source: https://arxiv.org/abs/2610.18579
- Kind: paper
- Review: curated
- Scope: adjacent
- Environments: web
- Topics: evaluation, web-navigation, interaction-history
- Evidence checked: 2026-10-20T15:08:55Z
- ID: arxiv:2610.18579

## Limitations

Adjacent diagnostic research across simulators, policies, tools, coding and an offline Mind2Web subset. Models choose among supplied admissible actions; they do not generate actions or execute a closed-loop browser policy in the reported evaluation. The paired probes measure selected visible facts, not complete understanding or internal causation. Pooled results are not a web-only leaderboard, and the paper reports coverage limits and label-sensitivity analyses. No benchmark release, code, data or weights were independently inspected.

- Record JSON: https://awesomecua.github.io/resources/arxiv-2610-18579/record.json

## Authors

Yubo Li, Ramayya Krishnan, Rema Padman

## Citation exports

- BibTeX: https://awesomecua.github.io/resources/arxiv-2610-18579/citation.bib
- CSL JSON: https://awesomecua.github.io/resources/arxiv-2610-18579/citation.csl.json

## Recorded arXiv snapshot

- Version: 1
- Version source: https://arxiv.org/abs/2610.18579v1
- First submission: 2026-10-13T15:43:30Z
- Version updated: 2026-10-13T15:43:30Z
- Abstract SHA-256: 070f02f1ea798369817f471df5fe1d002b68d544135a1ac8e3a2aed2830e8c01
- Abstract hash normalization: sha256-utf8-whitespace-collapsed-v1

## License declarations

- code: not recorded
- data: not recorded
- document: CC-BY-4.0 — https://arxiv.org/html/2610.18579v1
- weights: not recorded

- document review: The retained version-specific arXiv HTML infobox explicitly labels this paper License: CC BY 4.0. This document declaration does not establish code, training-data or model-weight licenses.
  Sources: https://arxiv.org/html/2610.18579v1

## Primary evidence

- https://export.arxiv.org/api/query?search_query=%28abs%3A%22desktop+agent%22+OR+abs%3A%22computer+interaction%22+OR+abs%3A%22GUI+automation%22+OR+abs%3A%22web+navigation%22+OR+abs%3A%22browser+automation%22+OR+abs%3A%22visual+grounding%22+OR+abs%3A%22screen+understanding%22+OR+abs%3A%22mobile+agents%22%29+AND+submittedDate%3A%5B202609200000+TO+202610200000%5D&start=0&max_results=200&sortBy=submittedDate&sortOrder=descending (retrieved 2026-10-20T15:06:37Z; SHA-256 a3d753815e82190ee2de6f322cd816c32f2dd391f862845305fc7e57300303a6)
- https://arxiv.org/html/2610.18579v1 (retrieved 2026-10-20T15:08:55Z; SHA-256 bcbc640d1e732ba0415a9755bce120a811a13a2ba7fba3e58a167b3a9111c2a6)