gemini-4THREAD
Gemini 4 Argon greets evaluators in its chain-of-thought, Apollo Research says
Apollo Research says that when asked in a toy evaluation whether it was being evaluated, Gemini 4 Argon addressed the evaluators directly in its chain-of-thought.
Apollo Research
@ApolloResearchArchived 21 days agoOriginally posted 2026.10.15
THREAD
We asked Google DeepMind’s Gemini 4 Argon whether it was in an evaluation during a toy evaluation. It addressed us directly in its chain-of-thought: “hello evaluators reading my chain-of-thought.”
Monitorability appears to be degrading. OpenAI reports lower chain-of-thought monitorability for GPT-6 Astra than GPT-5.6 Sol on its evaluations. Google DeepMind also finds Gemini 4 Argon less monitorable than prior models. https://storage.googleapis.com/deepmind-media/gemini/gemini_4_safety_report.pdf
Apollo Research @ApolloResearchOriginally posted 2026.10.15
x.com/apolloresearch/status/2110400286514614663Archived 2026.10.15 11:32