When an AI writes out its steps, is it unlocking reasoning it already has, or building that reasoning as it goes?
Does chain-of-thought trigger latent reasoning or create it?
This explores whether writing out step-by-step reasoning unlocks an ability the model already has, or whether the written steps are what build the reasoning.
This explores whether chain-of-thought (CoT) unlocks an ability a model already has or builds the reasoning through the steps it writes. The corpus leans toward trigger at the moment of use, and toward create only during training. The two get blurred together a lot.
The strongest evidence for trigger is that reasoning can be switched on with no chain of thought at all. Steering a single reasoning feature inside the model matches or beats CoT prompting across six model families. That mode turns on early in generation and overrides surface instructions (Can we trigger reasoning without explicit chain-of-thought prompts?). One broader reading is that reasoning happens in the model's hidden-state trajectory, and the visible text is only a partial window onto it (Where does LLM reasoning actually happen during generation?). Chain of Draft points the same way. It keeps accuracy while using 7.6% of the tokens, because the removed 92.4% was style and documentation rather than computation (Can minimal reasoning chains match full explanations?).
What CoT triggers also looks more like a pattern than like logic. It appears to constrain the model into reproducing familiar reasoning shapes, and it degrades predictably when the problem moves away from the training distribution (Does chain-of-thought reasoning reveal genuine inference or pattern matching?). Training format shapes strategy about 7.5 times more than domain does, and even invalid CoT demonstrations work about as well as valid ones (What makes chain-of-thought reasoning actually work?). The trigger can also misfire. For simple questions, answering directly beats step-by-step, because CoT only works when the question's meaning has already flowed into the prompt (Why do some questions perform better without step-by-step reasoning?). Accuracy also falls off with chains that run too long, and more capable models prefer shorter ones (Why does chain of thought accuracy eventually decline with length?). Both fit the idea that the words cue a capability rather than supply one.
Creation does happen, but in training rather than in the prompt. Treating CoT as an exploratory action during pretraining, rewarded by how much it improves prediction, lifted math and science benchmarks by about 19 percent (Can chain-of-thought reasoning be learned during pretraining itself?). Reinforcement learning (RL) can turn the same extended-thinking mechanism from self-doubt that hurts accuracy into useful gap analysis (Does extended thinking help or hurt model reasoning?). Building reasoning directly in latent space is hard. Outcome supervision alone starves the gradients and lets the latent space drift, so it needs dense supervision along the trajectory (Why does latent chain-of-thought fail so easily in training?). Even once it works, continuous thoughts need extra machinery, such as normalizing flows, to be sampled and scored the way text can (Can continuous thoughts have tractable likelihoods for sampling and scoring?).
The best short answer is that training creates the capability and CoT is one handle for reaching it. The written steps look more like a cue and an interface than the engine. The corpus doesn't settle whether long chains ever do real computation beyond cueing. The evidence here mostly shows how much of a chain is not computation.
Sources 11 notes
SAE-identified reasoning features can be directly steered to match or exceed chain-of-thought performance across six model families. This reasoning mode activates early in generation and overrides surface-level instructions, suggesting latent reasoning is a fundamental capability independent of explicit prompting.
Evidence from CoT faithfulness tests, feature steering, and layer analysis suggests latent-state dynamics drive reasoning, while surface chain-of-thought serves as a partial interface. Hidden reasoning processes should be the default focus of study.
Chain of Draft achieves equivalent accuracy to standard chain-of-thought on arithmetic, symbolic, and commonsense tasks while using only 7.6% of tokens. The 92.4% of removed tokens served style and documentation, not computation.
CoT works by constraining models to reproduce familiar reasoning patterns from training, not by enabling novel symbolic reasoning. Performance degrades predictably under distribution shifts—the signature of imitation rather than capability emergence.
Research shows training format shapes reasoning strategy 7.5× more than domain, demo position swings accuracy 20%, and invalid CoT prompts work as well as valid ones. CoT is pattern-guided generation, not formal logic.
Show all 11 sources
Saliency analysis reveals that CoT prompting fails when question information doesn't aggregate into the prompt structure before reasoning begins. For simple questions, direct question-to-answer flow outperforms step-by-step reasoning, showing the optimal prompt depends on question type, not just task category.
Task accuracy peaks at intermediate CoT length, with optimal length increasing alongside task difficulty but decreasing with model capability. RL training naturally gravitates toward shorter chains as models improve, revealing that simplicity emerges from reward signals rather than explicit training.
RLP treats CoT as exploratory action during pretraining, using log-likelihood improvement as verifier-free reward. Applied to Qwen3-1.7B and Nemotron-Nano-12B, the method improves math and science benchmarks substantially, suggesting reasoning can be planted earlier in training.
Vanilla models use thinking mode counterproductively, inducing self-doubt that degrades performance. RL training reverses this, transforming the same mechanism into beneficial gap analysis. Training mediates reasoning quality, not just quantity.
Outcome supervision alone causes gradient attenuation along latent steps and lets the latent space wander without semantic grounding. Robust latent reasoning requires both dense trajectory supervision and space supervision that preserves geometric structure rather than compressing it.
NF-CoT models continuous thoughts as an autoregressive normalizing flow inside the LLM's causal stream, recovering exact likelihood, probabilistic sampling, and KV-cache compatibility. This enables policy-gradient refinement and trajectory scoring on non-verbal reasoning, matching the tractability of textual CoT.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Hierarchical Reasoning Model
- Reasoning Beyond Chain-of-Thought: A Latent Computational Mode in Large Language Models
- When More is Less: Understanding Chain-of-Thought Length in LLMs
- Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens
- Break the Chain: Large Language Models Can be Shortcut Reasoners
- CoT is Not True Reasoning, It Is Just a Tight Constraint to Imitate: A Theory Perspective
- Eliciting Reasoning in Language Models with Cognitive Tools
- Base Models Know How to Reason, Thinking Models Learn When