Distinct internal patterns for reasoning steps

A new study finds that models exhibit distinct internal patterns for different written reasoning steps—calculation, formula retrieval, and deduction—with these patterns strongest in middle layers. This shows models internally process more than their visible chain-of-thought, affecting interpretability and AI safety analysis.

• OpenAI Has Contractors Reviewing ChatGPT Chats2026-09-14
• Two-Year Study Finds Banning AI Hurts Students2026-09-13
• OpenAI Questions Legality of an Industry Slowdown2026-09-11
• Analysis: Claude Fable 5.1 writes more plainly2026-09-10
• Hugging Face launches ML Intern to run ML experiments via chat2026-09-09