Artificial Intelligence Learns to Reason
science.org
science.org
Large Reasoning Models (LRMs) are a new type of AI system designed to perform specific types of reasoning. They generate “chains of thought” to solve problems, mimicking human reasoning processes. While LRMs have shown impressive performance on benchmarks, their robustness, trustworthiness, and interpretability remain uncertain, raising questions about their ability to handle real-world problems.
if in math formulae, you altered a variable name, solving suddenly failed.
Could the example of "solve count of siblings" from the paper be implicit connections in the model, instead of reasoning?
I don't think I'd have expected differently even in 1990.
False. Julie could be the name of a boy.
> 2. Julia also has one brother, named Martin.
OK
> 3. Altogether, there are four siblings: three girls and one boy (Martin).
Where did you identify the third girl?
> 4. From Martin’s perspective, his sisters are all three of the girls (Julia and her two sisters).
Here's where the answer comes from (Julia and her two sisters).
More directly: "How many sisters does her brother Martin have?"
> 5. Therefore, Martin has three sisters.
OK.
So the reasoning might have come to the "right" answer but the way it arrived at the answer was incorrect.