Machine Learning Street Talk (MLST) avatar

Dexa/Machine Learning Street Talk (MLST)

Learn more
Published Jul 29, 2024

Prof. Subbarao Kambhampati - LLMs don't reason, they memorize (ICML2024 2/13)

Prof. Subbarao Kambhampati discusses the limitations of large language models, emphasizing their tendency to memorize rather than reason. He proposes the LLM Modulo Framework that enhances reasoning by integrating LLMs with external systems, combining creativity with logical verification for greater reliability in complex tasks.
Episode Highlights
Machine Learning Street Talk (MLST) logo

Topics covered

Popular Clips

Episode Highlights

  • Creative Power

    Large Language Models (LLMs) excel in generating creative ideas, leveraging their extensive training data to make connections that might elude the average person. highlights their ability to create analogies and generate hypotheses, akin to the creative leaps made by mathematicians like Ramanujan 1 2. However, he emphasizes the need for verification of these creative outputs to ensure their validity.

    The only thing that civilization needs is that when you get the idea, you check whether the idea works.

    ---

    This underscores the importance of combining LLMs' creative capabilities with human or external verification systems to ensure robustness 3.

       

    Reasoning Limits

    Despite their creative prowess, LLMs struggle with reasoning tasks that require guarantees of correctness. points out that while LLMs can generate stylistically impressive content, they lack the ability to ensure factual accuracy 4. This limitation is evident in tasks like planning or solving procedural problems, where LLMs fail to generalize solutions beyond simple examples 5.

    They are not going to give you any guarantees about instance level correctness.

    ---

    Misconceptions about LLMs' capabilities often arise from incorrect assumptions about computational complexity, leading to overestimations of their reasoning abilities 6.

       

    Interdisciplinary Use

    LLMs hold potential for interdisciplinary applications by leveraging their creative strengths while addressing reasoning weaknesses. suggests integrating LLMs with external verification systems to enhance their utility in fields like code generation, where creative outputs need validation 7. This approach aligns with the concept of using LLMs as part of a broader system that includes human oversight and external checks.

    We are in this interesting world where we are actually developing this huge, large models, and you are poking it to see what they are trying to do.

    ---

    Such integration can mitigate the risks associated with emergent abilities and ensure that LLMs contribute effectively to various domains 8.

Related Episodes