Report
Selecting one of five runs is claimed to lift Fable 5.1 from 58% to 70% on Terminal-Bench 4.0
A user says they hid an evaluation result and asked another agent whether the run worked.
TLDR
A user says agent evaluations grade thousands of runs, then use each only once for a score. Their team hid a run’s result and asked another agent whether it worked. They say models are bad judges but good choosers: selecting one of five runs lifts Fable 5.1 from 58% to 70% on Terminal-Bench 4.0.
Combined views
—
2 Sources, first seen ago
