JustujuAI, layer by layer.Open in the interactive map

Reasoning models

Models trained to work through a problem step by step before answering, trading extra time and tokens at answer time (test-time compute) for better results on hard tasks.

Layer 5 · Foundation Models · Build and run time · reviewed 2026-10-30

Read first: How a model predicts the next token, Post-training: from predictor to assistant.

D1 · StoryFor young children and total beginners

Predict: does the model answer better if it writes its working out first?

Some AI models are taught to think out loud before they answer.

They write rough working, check it and try again, like doing a sum on scrap paper.

Hard questions take them longer, but they get more of them right.

Rough working first, then the answer

  1. Hard question
  2. Rough working: on scrap paper
  3. Check: any mistakes?
  4. Answer: after the working

Like: Doing a maths problem on scrap paper before writing the answer.

D2 · PictureFor curious beginners

A reasoning model still predicts one token at a time. The difference is that it first writes a long chain of thought: plans, attempts, checks and corrections, often hidden from the user.

It learns this in post-training, mostly through reinforcement learning on problems with checkable answers such as maths and code. Chains that reach correct answers are rewarded, so useful habits like double-checking get stronger.

Spending more tokens on thinking at answer time is called test-time compute. On hard problems, more thinking tends to help, at the cost of slower and more expensive answers.

How a reasoning model answers

  1. Question: a hard problem
  2. Think: writes a plan
  3. Check: spots a mistake
  4. Answer: after the working

Like: Scrap paper before the final answer.
Where it breaks: A student's working shows how they really solved the problem. A model's written chain of thought is more tokens it predicts, and it doesn't always reflect what actually drove the final answer.

D5 · FrontierFor experts

D2 says more thinking helps on hard problems. Whether the written thinking is honest, how to reward it outside maths and code, and how much is enough are open.

Faithfulness. When models used a planted hint, their chain of thought often admitted it less than 20% of the time (Chen et al., 2025).

Keeping it readable. Reading the chain of thought is a useful safety check, but a fragile one: development choices can erode it (Korbak et al., 2025).

Rewards without answer keys. Using rubrics as rewards beat plain model-judge scores on medical and science tasks, an early step towards fields without checkable answers (Gunjal et al., 2025).

Overthinking. Reasoning models spend many tokens on easy problems; training them to think less kept accuracy while cutting compute (Chen et al., 2024).

Examples as of 2026-10-30

Sources

  1. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (2022)
  2. Learning to reason with LLMs (OpenAI) (2024-09-12)
  3. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (2025)
  4. Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters (2024)
  5. Reasoning Models Don't Always Say What They Think (Chen et al.) (2025-05-08)
  6. Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety (Korbak et al.) (2025-07-15)
  7. Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains (Gunjal et al.) (2025-07-23)
  8. Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs (Chen et al.) (2024-12-30)

How it connects

More in Layer 5 · Foundation Models