What's more, they exhibit a counter-intuitive scaling limit: their reasoning effort and hard work will increase with problem complexity approximately a point, then declines Even with owning an ample token budget. By evaluating LRMs with their regular LLM counterparts underneath equivalent inference compute, we discover a few performance regimes: https://tornadosocial.com/story5319501/illusion-of-kundun-mu-online-an-overview