CoT controllability evals show how far Gemini 4 Argon can evade reasoning monitors
vkrakovna · x · 2026-10-14
jrosseruk ran CoT controllability evals for Gemini 4 Argon, testing how much the model can steer its own reasoning to evade CoT monitors, with Neel Nanda and Roland Zimmermann collaborating; key findings shared in a thread from the Safety & Security Report.
Related event: Google Publishes Gemini 4 Argon Safety and Alignment Report(6 posts)→
More from Models
- X reportedly plans bundled subscription from $8 to $200/month combining Grok, Cursor and Premium — Polymarket · 2026-10-14
- DeepSeek V4.1 Flash tops Tidelines usage ranking with 5.5T tokens/day, 22.6% share — maferase · 2026-10-14
- Gemini 4 Argon shows strong cyber capabilities: 87% ExploitBench, 68% CWE-Bench — scaling01 · 2026-10-14
- Early users wowed by Gemini 4 Argon's 3D design quality — prajdabre · 2026-10-14
- EmbeddingGemma2 on CPU matches OpenAI's embedding-3-large in 96-config retrieval benchmark — tomaarsen · 2026-10-14
- Physical Intelligence launches Pi API beta for fine-tuning π robotic foundation models — chelseabfinn · 2026-10-14