LLMs’ “simulated reasoning” abilities are a “brittle mirage,” researchers find
arstechnica.com
Chain-of-thought AI “degrades significantly” when asked to generalize beyond training.