4 ms·

LLMs struggle with math reasoning, because they can't conjecture

• jqpabc123 1y ago
  LLMs struggle with reasoning --- period, the end. They are fundamentally probabilistic, not deterministic. This makes them inherently unreliable for any application where deterministic logic is required --- like math. Don't take my word for it, just ask your favorite LLM. --- "Can you guarantee your results are reliable?". While I strive to provide accurate information and results based on the data I have, I can't guarantee absolute accuracy in every case. --- "Do you hallucinate?" Yes, I can "hallucinate" in the sense that I might sometimes generate information that is factually incorrect, misleading, or doesn't fully align with reality.
  • trehcrob 1y ago
    It's not just that LLMs are probabilistic, it's that the illusion of reasoning in English doesn't transfer to formal systems. I find it quite interesting how far LLM reasoning can be pushed in for example code generation. But there is a massive difference between plausible argumentation and provably correct statements.
    • jqpabc123 1y ago
      it's that the illusion of reasoning in English doesn't transfer to formal systems. This is because there is very little reasoning --- it's an "illusion". Data retrieval is not "reasoning". A person with a photographic memory and instant recall is not necessarily a "genius" --- though it may appear that way to some.