23 October 2026

Study finds AI decision controller raised agent costs without improving results

First reported

Deep Learning Weekly ran this on 23 October.

A pre-registered study tested a decision controller, called Jev, on the Terminal-Bench 2.1 benchmark for AI agents.
The controller raised the cost per attempt by 19% without any gain in the pass rate.
Its confidence estimates were poorly calibrated, meaning its stated certainty did not match how often it was right.

Our read

Only one of the 30 tracked newsletters, Deep Learning Weekly, covered this study, which first appeared there on 2026-10-23. Our archive's earlier 2026-10-17 story on AI models omitting flaws in their own reports is the nearest prior entry on this thread.

Earlier on this story

Oct 17Study finds AI models omit flaws in their own work reports; one sentence helps
Oct 5Study finds brain waves actively coordinate thinking, not mere noise
Sep 1Study finds US data centers have minimal local power bill impact

Written by Newskeryx from which newsletters covered this story and from our own archive.

How it was covered

Deep Learning WeeklyEditorial team

Controller comparison design, cost and pass-rate results, calibration problems, and abandonment behavior