Implicit chain-of-thought · A reading list
Can language models reason without writing every intermediate step?
An explicit chain of thought writes out intermediate steps. Implicit chain-of-thought methods try to do some or all of that work without putting each step into words. This research is also discussed as latent reasoning or latent chain-of-thought.
Yuntian Deng and colleagues first proposed implicit chain-of-thought reasoning in 2023: training language models to reason through continuous hidden states rather than write out explicit intermediate steps. The paper's distillation method is one implementation of that proposal. Their 2024 work introduced step-by-step internalization.
Try an example: Multiplication · Math word problems
How do these methods work?
Some learn to skip written steps. Others add computation between words or reuse model layers. The list also covers related generation methods and studies of when these approaches work or fail.
Internalized reasoning
Learn from worked solutions, then answer with fewer or no written intermediate steps.
Explanation and papers →Thinking tokens
Add pauses or special tokens that give the model more computation without spelling out a solution.
Explanation and papers →Continuous-state reasoning
Pass vectors between reasoning steps instead of choosing a text token at every step.
Explanation and papers →Recurrent reasoning
Reuse model layers or carry information in memory to spend more computation on a problem.
Explanation and papers →Alternative generation methods
Change how steps are generated—for example, by refining several together. Some methods still use written reasoning.
Explanation and papers →Analysis and evaluation
Test what these methods can do, how they work, and when they fail.
Explanation and papers →About this reading list
This reading list grew out of the Latent & Implicit Thinking workshop at ICLR 2026. You can still find the workshop program and event information in the archive.
Thanks to the Awesome-Latent-CoT bibliography and its associated survey, which helped us find many of these papers.
Suggest a paper or correction.
For text readers and research tools: /llms.txt (reading guide) · /data/navigation.json (page and download URLs).
