Thore Graepel questions how large language models reason
What happened
On October 2, MIT Technology Review reported Thore Graepel's doubts about the reasoning ability of large language models. He argues that a model's chain of thought still relies on next-token prediction and lacks the separate search and reasoning mechanism AlphaGo had. Chatbots usually have no explicit, persistent, inspectable knowledge state, and knowledge is not clearly separated from reasoning, so the generated chain of thought may not match the path that actually produced the answer. He proposes borrowing AlphaGo's architecture: keep an updatable knowledge state, and have a separate mechanism judge from evidence whether each step reduces uncertainty, so the way a conclusion forms can be audited.
Written by AI from the coverage · updated 1 hour ago
Coverage
Follow the reports to see the story from different sides.
- MIT Technology Review · AIDon’t be fooled—LLMs don’t reason
Thore Graepel 主张，大语言模型的思维链仍依赖下一 token 预测，缺少 AlphaGo 那样独立的搜索与推理机制。他指出，聊天机器人通常缺乏显式、持久且可检查的知识状态，知识与推理没有清晰分离，生成的思维链也可能与实际得出答案的路径不一致。他提出借鉴 AlphaGo 的架构，维护可更新的知识状态，由独立机制依据证据评估每一步是否减少不确定性，使结论形成过程可以审计。
Heat over time
Not enough continuous observations to draw a trend yet.