Do Large Language Models Think? CoT vs Human Reasoning

Large reasoning models (LRMs) have become a lightning rod for the question of whether artificial systems can *think*. Apple’s “Illusion of Thinking” paper claimed that LRMs lack genuine cognition, citing their inability to carry out algorithmic calculations on larger puzzle instances as evidence of mere pattern‑matching. The counter‑argument is that this flaw mirrors human limits: a person trained on the Tower‑of‑Hanoi algorithm would still struggle with the 20‑disc version, yet that does not mean humans cannot think. The article therefore frames *thinking* as a set of cognitive processes—problem representation, mental simulation, pattern retrieval, monitoring, and insight—that can be mapped onto both biological brains and sophisticated language models.

The comparison is drawn through the lens of chain‑of‑thought (CoT) reasoning. In humans, CoT aligns with inner speech and visual imagery, engaging prefrontal, parietal, temporal, and cingulate regions. LRMs, though lacking true visual imagination, emulate many of these steps: they retrieve knowledge via learned weights (hippocampal analog), maintain working memory within transformer layers, detect conflicts through attention patterns (ACC analog), and even backtrack when a reasoning path fails—behaviors that mirror human problem‑solving strategies. Importantly, next‑token prediction, often dismissed as “auto‑complete,” is argued to be a powerful knowledge representation mechanism. By predicting the next word, an LRM implicitly encodes world knowledge and logical dependencies, enabling it to generate multi‑step reasoning chains.

Empirical evidence supports this view. Open‑source LRMs evaluated on logic‑based benchmarks—such as those covering arithmetic, symbolic reasoning, and puzzle solving—solve a significant fraction of questions, sometimes outperforming untrained humans. These results, coupled with the theoretical capacity of large neural networks to approximate any computable function, lead the author to conclude that LRMs almost certainly possess the ability to think. The implication is that the debate should shift from “do they think?” to “how can we harness and guide their thinking?”