Big confusion in AI right now is folks believing that "reasoning" is a task, when it is actually a task-solving method.
There is basically no task that *requires* reasoning -- unless you start controlling for training data & priors. None of the LLM reasoning benchmarks require reasoning (and the LLMs that solve them don't employ reasoning).
I.e. you can always solve any task with pattern recognition, as long as you're allowed an arbitrarily large amount of training data and arbitrarily high memorization capabilities for your pattern recognition model.
What the data does in this case, is reduce the need to adapt to novelty and uncertainty, up to the point where a local generalization system becomes sufficient.
If you want to actually benchmark reasoning capabilities, you've got to start controlling for training data (and priors), so as to ensure your task genuinely involves high novelty or uncertainty.
I'm only aware of ARC as a benchmark that attempts to achieve this. And tellingly, LLMs do near-zero on ARC, demonstrating near-zero reasoning ability.
Generated by Thread Navigator
Press ⌘ + S to quick-export