NVIDIA Mid-Harness raises TerminalBench-Lite Pass@1 from 50.0% to 68.0%
An NVIDIA paper on test-time compute for terminal agents says a GPT-5.6 Sol verifier choosing among eight sampled shell commands lifts TerminalBench-Lite Pass@1 from 50.0% to 68.0%. Extra samples add almost nothing when the verifier is weak.