Terminal coding agents catch 61% of their wrong answers and fix half of those

TL;DR

• Across ten terminal agents on TerminalBench 2.1, the authors find that agents check their work 99.53% of the time but detect only 61.43% of their wrong candidates and repair only 49.36% of the errors they detect.
• Their method, Student-Conditioned Verification Distillation, has a stronger model continue from the student's own candidate and teaches the student that checking; Pass@1 rises 9.74 to 16.85 points over the base models.
• On SWE-bench Verified the method holds or improves while ordinary distillation drops up to 25.87 points. The paper is the only source.