博士论文 | 理解与改进大语言模型推理：从反转诅咒到连续思维链

TL;DR - This UC Berkeley doctoral thesis analyzes LLM reasoning reliability and efficiency, linking failures such as the reversal curse and knowledge-combination hallucinations with approaches for latent-space reasoning. It argues continuous chain-of-thought can represent multiple candidate paths simultaneously, potentially reducing reliance on costly textual reasoning traces.

• Training on a relation in one direction may not teach its inverse because autoregressive models can memorize sequence-level patterns rather than abstract relational rules.
• Out-of-context reasoning provides a shared framework for useful knowledge composition and hallucinations caused by incorrect associations after knowledge injection.
• Continuous CoT maintains superpositions of candidate solutions in vector space, enabling implicit parallel search without prematurely committing to one textual path.
• Theory and ProsQA graph-reasoning experiments indicate continuous reasoning can solve reachability tasks efficiently, with the superposition mechanism emerging during training.