$ cat briefs/daily/2026-10-17.md
2026-10-17
October 17, 2026 (Sat)
1 story · 1 paper pick · no wiki changes in this reconstruction
This archive was reconstructed on **2026-10-18 KST** after the original automation failed. Its original morning cutoff is **2026-10-17 08:00 Asia/Seoul**, equivalent to **2026-10-16 23:00 UTC**. It was not published or emailed on the missed date; this reconstruction is archive-only, with no subscriber delivery. **Provenance and limits:** the research below uses existing snapshots captured on **2026-10-18**, checked against the linked **v1** abstracts. Submission timestamps alone are not publication timestamps: both papers also appear in arXiv's **2026-10-16** announcement listings, before the cutoff. Full papers and code were not reviewed; this is a limited reconstruction, not a complete record of that morning. Wiki links provide present-day navigation, not evidence that those pages existed at the cutoff. ([OpenGCUA listing](https://arxiv.org/list/cs.CV/recent?skip=0&show=2000)) ([FDE-Gym listing](https://arxiv.org/list/cs.AI/recent?skip=0&show=2000))
Top Stories
1. OpenGCUA makes interface choice part of the agent
The authors combine code execution with visual interaction and move selected observation and task-state decisions from fixed harness rules to the model. Across 8 models, the full configuration raises OSWorld-v2 mean partial score by +19.6 points and reduces reported steps by 53.0% against batched GUI-only execution. Partial credit is not task completion, and step counts do not establish total cost.
Why it matters: the result makes the agent's interface and decision boundaries explicit parts of evaluation. → OpenGCUA (snapshot) (arXiv v1)
Paper Picks
FDE-Gym — discover requirements, then deliver. In simulated customer environments, the strongest realistic-setting configuration resolves 14 of 95 tasks and averages 0.56 on a score measuring normalized customer-metric improvement with interaction penalties. Resolution requires at least 0.8 with no incident. Read it for evaluating unstated constraints and successful execution together; these are simulated delivery results, not measured production impact. → FDE-Gym (snapshot) (arXiv v1)
Watch
No additional independent signal met this reconstruction's cutoff-verification standard. An empty watch section does not establish that the morning was quiet.
New in Wiki
No pages were created by this archive run. The linked research notes were originally created on 2026-10-18, as their frontmatter records; they are not additions from the missed 2026-10-17 morning. → OpenGCUA · FDE-Gym. (OpenGCUA note) (FDE-Gym note)
Updates
No existing wiki pages were edited by this reconstruction. The retained log has no 2026-10-17 entry, so original work on that date is not asserted. (log)