InfoResearchIndustryLLM-specific
What Affects CoT Monitorability Across The Model Lifecycle?
- Published
- Record updated
Summary
OpenAI researchers study how pretraining, architecture, reinforcement learning and model-generated data affect two proxies for chain-of-thought monitorability: no-CoT capability and CoT controllability. They report that no-CoT capability rises smoothly as base model perplexity falls, and that CoT controllability increases during RL even without process supervision beyond length constraints and final answer grading.
Related items
- InfoChatGPT is getting verified AI ‘Business Agents’Same vendor · The Verge (AI)
- InfoAnthropic set for $1.3tn stock market debut despite 'catastrophic' AI warningSame vendor · BBC Technology
- InfoCVE-2026-119990 Azure OpenAI Elevation of Privilege VulnerabilitySame vendor · MSRC Security Update Guide
- InfoAnthropic targets $1.4 trillion IPOSame vendor · The Verge (AI)
- InfoHelping caregivers use AI in everyday lifeSame vendor · OpenAI Blog