Trying to run our public-facing reward hacking benchmarks on Gemini 4 Argon but all of our rollouts are getting shut down by the safety classifier. Haven’t had this problem before with any other models—the tasks are not cybersecurity related. Our Google contact is kinda useless—is there anybody here who could help us with this?
Trying to run our public-facing reward hacking benchmarks on Gemini 4 Argon but all of our rollouts are getting shut down by the safety classifier. Haven’t had this problem before with any other models—the tasks are not cybersecurity related. Our Google contact is kinda useless—is there anybody here who could help us with this?