Skip to content
LessWrong· Simon Mendelsohn·Indexed by this site · Originally published· Opinion →SelectedAttention40

"About a quarter contain defects": the ML R&D benchmark in the Gemini 4 safety report

"About a quarter of them contained bugs": Gemini 4's ML R&D benchmark

AI summary

LessWrong author Simon Mendelsohn analyzed the Gemini 4 Argon safety report and pointed out that in GRB (Google RE-Bench, 74 tasks), the internal benchmark Google used to rule out the ML R&D automation threshold, about a quarter of the tasks have defects and no submission can pass, and several other tasks are scored according to details not stated in the task description. The report sets a 90% pass rate as the conservative threshold and says that even if known defective tasks are counted as solved, Argon's pass@1 is still below the threshold, but it does not publish specific numbers or score-point estimates. The report says Argon solved about half of the tasks in completed runs; these runs skew toward short tasks and may overestimate its true pass rate, and on the same batch of completed trajectories it performs comparably to Gemini 3.7 Flash, whose pass@1 on GRB is 27%; even if all incomplete trajectories succeeded, Argon would still be below the threshold.

Why it was selected

After checking the ML R&D benchmark in the Gemini 4 safety report, the author points out that about a quarter of the tasks cannot be passed at all, which determines how persuasive its threshold-exclusion conclusion can be.

Related papers
Read originallesswrong.com