A new study from the Massachusetts Institute of Technology (MIT) has concluded that large language models (LLMs) do not exhibit human-like behavior. This discrepancy arises because these models are trained on data that may not align with human generalization functions or expectations. Researchers highlight that LLMs, such as GPT-4, do not show reasoning that emerges from distributional learning. Instead, their reasoning capabilities are limited to memorization and interpolative retrieval. Human supervision and symbolic methods in post-training are necessary to improve their reasoning abilities. The study, published on July 23, 2024, refutes the naive idea that GPT-like models think like humans, emphasizing that they do not even think in the way they describe their thinking. The Llama 3 paper provides a description of what is needed to get an LLM to reason more effectively.
Sources
martin_casado
"... pretty clearly refute the naive idea of what GPT-like models do. These models are not thinking just like we do. They are not even thinking in the way they are describing their thinking." Excellent read on LLM reasoning reducing to fuzzy matching. https://t.co/1NfumF2gVu
François Chollet The question of whether LLMs can reason is, in many ways, the wrong question. The more interesting question is whether they are limited to memorization / interpolative retrieval, or whether they can adapt to novelty beyond what they know. (They can't, at least until you start…
Tal LinzenOne reason I'm excited about the Llama 3 paper is we finally have a description of what you need to do to get an LLM to reason sort-of well (human supervision, symbolic methods in post-training). GPT-4 and co. do not show that reasoning "emerges from distributional learning". https://t.co/t8NGPxQuq4