CoT controllability rises during RL training and tracks no-CoT capability across model generations
tomekkorbak · x · 2026-09-04
Researcher tomekkorbak shares root-cause analysis from recent weeks: CoT controllability increases over the course of RL training, which was not the case for previous models, and it is strongly correlated with no-CoT capabilities across several generations of models.
Related event: DeepMind Researcher Warns CoT Monitorability Is Declining(4 posts)→
More from Research
- Python dicts and sets can hit quadratic time: the O(1) assumption breaks down — lemire · 2026-09-04
- Mathematician gives CMU talk on OpenAI's proof of a non-sofic group — littmath · 2026-09-04
- Millière vs Mitchell: can the intentional stance make an AI bot a genuine believer? — raphaelmilliere · 2026-09-04
- Applying the Expected Value Framework to Get ROI from Data Science — mdancho84 · 2026-09-04
- Treating Machine Learning Predictions as Probability Estimates — mdancho84 · 2026-09-04
- CRISP Data Mining Process as the Foundation for Business Data Science — mdancho84 · 2026-09-04