A newly printed apple tool learning learn owns checked the tale prevalent close AI “logic”, such like Openai’s O1 , and Cloud’s Thinking variants such like large words models, saying of easy boundaries the one propose the one those systems are not really arguing.
For learnInstead of using normal mathematics benchmarks,; But touch hurt from data creating not clean, Apple’s researchers formulated a commanded puzzle air close owning in the Tower of Hanoi , and River Crossing. Also, like spoke to the researchers, two the; But answer , and inside arguments let the right learn of two in not same hardness levels.
The results are striking to speak in least; But all checked arguments models-in which O 3-Munes, Dipsec-R1, and Cloud 3.7 Sonnets-part hardness lived through a lot of in it being right past threshold, and fell to zero winning speed despite enough computational resources. Counters, models really reduce their thinking try why problems become more hard, propose easy scaling boundaries instead than provide barriers.
Maybe the most hurting, flat. Researchers given a lot of in it answer algorithms, the models still failed in the same hardness points; But researchers speak the one thing indicates the one the limit is not in the plan of solving the trouble. Also, but the easy reasoning move is in doing.
The model too displayed discrepancies – to beat on problems needing 100+ moves. Also, failing on easy riddles needing just 1 tricks.
Learn highlights three not together doing regions: normal models shocking path do models on low hardness, logic models show help on center hardness, and two approaches fail a lot of in ity in tall hardness; But the learn of the researchers of the word fight marks displayed the can’t work “ovethinking” plan, where the model found the right answer. Also, but ruined the computational budget searching for wrong options.
The tech-hom of Apple’s findings is the one the now “logic” models rely on the matching of the refined plan instead than the real logic capabilities; But here suggests the one llms don’t argue like humans, remove easy problems , and think less for hard persons.
The when of printing is notable, wwDC owns surfaced a few days earlier 2025, where Apple hopes the one Apple is expected to limit his focus on AI in help of new software designs , and features.Bloomberg,
Source link
