Moreover, they show a counter-intuitive scaling limit: their reasoning work improves with challenge complexity approximately a point, then declines despite possessing an sufficient token budget. By evaluating LRMs with their normal LLM counterparts under equivalent inference compute, we detect 3 overall performance regimes: (one) minimal-complexity tasks wherever regular models https://illusion-of-kundun-mu-onl77654.csublogs.com/42963279/not-known-factual-statements-about-illusion-of-kundun-mu-online