In this guide
Reasoning models use intermediate steps to work towards an answer, and training can make those steps more useful. At response time, a system can spend computation on a longer attempt, several attempts, or a search that scores alternatives. The benefit depends on the task and how that work is spent. A longer response alone guarantees neither a correct answer nor a complete explanation of how it was reached.
A shelf starts with 18 books. Six are removed, then two boxes containing four books each are added. How many books are on the shelf?
A worked solution is 18 − 6 = 12; 2 × 4 = 8; 12 + 8 = 20. Those intermediate results give the next step something to use. A mistaken solution might add only two books and finish at 14. Producing another attempt could help, but selecting between 14 and 20 still needs a useful check. For this example, the wording of the problem and the arithmetic establish that 20 is correct.
A chain of thought, or CoT, is a generated sequence of intermediate steps. A token is a chunk of text processed or produced by the model; generating more tokens requires more computation. Inference means running a trained model to produce a response. Test-time compute, also called inference-time compute, is the computational work spent on the current input.
Separate the prompt, training and answering
The five readings below examine changes at different stages. Keep them separate when an author says that a model has become better at reasoning.
- Examples in the prompt. The input includes worked solutions that demonstrate a pattern. For the shelf task, show another problem that multiplies boxes by books per box before adding the total.
- Training. Examples or rewards change the model's learned behavior. Reward correct solutions across many tasks so useful solving behavior becomes more likely.
- Inference. The system spends work generating, revising or choosing candidate solutions. Try another calculation, revise the first attempt or score competing answers.
Worked examples in a prompt do not by themselves retrain the model. Likewise, extra computation during one response is a separate choice from the computation used to train it. A verifier scores candidate answers or intermediate steps; its judgment can also be wrong.
Read the following sequence for a map of the methods, their experimental origins and the limits of what the written steps reveal. The first item is an author's research review. The later items report the research teams' own experiments.
1. Get a map: Lilian Weng, Why We Think
Author's research review · 1 May 2025
Why We Think brings together ways to spend more computation while answering. Open Lilian Weng's profile for the surrounding reading context, then use this review to distinguish methods before following their experiments.
Start with: Thinking in Tokens, followed by Branching and Editing. The useful distinction is between generating several candidates and revising an existing one. The approaches can be combined; asking for another look does not automatically produce a correction.
What you need: the token and inference definitions above. Leave the latent-variable equations and recurrent architectures for a later pass.
Keep the boundary: this is Weng's synthesis of other teams' work. Its analogy with slower human thinking motivates a question; it does not establish that a language model uses the same mechanism as a person.
Where extra work can go at answer time
These are different operations during inference. Generation and selection can be combined.
- Continue one attempt Generate a candidate
- What happens Produce intermediate steps before the answer within one ongoing response.
- What to check More generated steps do not by themselves establish correctness.
- Sample alternative attempts Generate candidates
- What happens Generate independent responses to the same problem, potentially exploring different approaches.
- What to check Choosing one final answer requires a selection rule; attempts may share the same mistake.
- Evaluate and select Assess candidates
- What happens Use a verifier's scores or an aggregation rule, such as answer agreement, to choose among candidates. Scoring may also guide an unfinished search.
- What to check A learned score or a majority answer can still be wrong.
Sequential revision is another option: generate a new answer using an earlier attempt as context. It can be combined with sampling and selection.
2. See the prompt experiment: Jason Wei and collaborators
Original research · first submitted 28 January 2022; linked revision 10 January 2023
In Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, Jason Wei and eight coauthors at Google Research examine prompts containing examples with intermediate steps.
Start with: the introduction and section 2, then read the limitations in section 6. Compare ordinary question–answer examples with examples that include the steps leading to the answer. The researchers elicited this behavior from existing models without fine-tuning them for the experiment.
What you need: the difference between an example supplied in the current input and training that changes the model.
Keep the boundary: this is few-shot prompting, using worked examples. It is not the complete training recipe for modern reasoning models. Generated steps can be wrong even when an answer looks convincing. Jason Wei's profile connects this coauthored work to his other research.
3. Ask where the extra computation goes: Charlie Snell and collaborators
Original research · 6 August 2024
Charlie Snell, Jaehoon Lee, Kelvin Xu and Aviral Kumar study the allocation question in Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters. Should a system generate fresh attempts, revise an answer or search with a verifier?
Start with: the introduction and sections 2–3. Follow the generation-and-verification setup before the equations. Sections 5–6 then compare search and revision.
What you need: a verifier scores candidates; a compute budget limits the work available. Several attempts and one longer attempt use that budget differently.
Keep the boundary: the experiments use particular models and mathematical problems. The useful allocation depends on problem difficulty relative to the model. Estimating that difficulty also costs computation, largely excluded from this analysis. The results do not establish a universal replacement for larger models.
4. Follow a training recipe: DeepSeek-R1
Technical report · DeepSeek-AI · 22 January 2025, version 1
The original DeepSeek-R1 report gives a documented training case. It distinguishes R1-Zero from the multi-stage R1 pipeline.
Start with: sections 2.1, 2.2.2 and 2.3, then the limitations in section 5. Skip the GRPO equation on the first pass.
What you need: supervised fine-tuning, or SFT, trains on examples; reinforcement learning, or RL, trains using rewards. Distillation here trains other models on answers generated by a teacher model.
Keep the boundary: R1-Zero starts from a pretrained base model, with rule-based rewards for correctness and formatting and no preliminary SFT. R1 adds supervised cold-start data and further training stages. “No SFT” therefore does not mean “trained from nothing.” This is one team's recipe. The version is pinned because the later report differs.
5. Test what the explanation reveals: Anthropic's faithfulness study
Research team's explanation · Anthropic Alignment Science · 3 April 2025
Reasoning models don't always say what they think explains experiments by Yanda Chen and collaborators. It addresses a different question from whether the final answer is correct: does the written reasoning acknowledge a cue that influenced the answer?
Start with: the full short explanation, including its conclusions. Researchers inserted hints into multiple-choice questions. When a hint changed the model's answer, they checked whether its CoT mentioned that influence. The linked research paper gives the fuller method.
What you need: faithfulness concerns how accurately an explanation represents the factors behind an answer. It is a separate property from correctness.
Keep the boundary: the experiments tested Claude 3.7 Sonnet and DeepSeek R1 on particular hints and artificial tasks. The hints could be used without CoT, so these tests do not settle how well monitoring works when a task depends on it. An omitted hint does not establish conscious deception. The findings limit what CoT monitoring can rule out; they do not make all monitoring useless.
What you can now distinguish
Return to the shelf problem. Checking the arithmetic establishes that 20 is correct. A polished explanation ending at 14 still fails that check. Even a correct answer with sensible steps would leave a further question: did those steps disclose every influence on the answer? The faithfulness study shows why the written explanation alone cannot settle that question.
When reading the next reasoning claim, locate its change: examples in the input, a training method, or computation used while answering. Then identify what was measured and which tasks the result covers. That lets you separate an effective method in a documented setting from a promise about every future answer.
For more authors and first materials, continue with the LLM research collection. If your question is whether a change helps your own application, use the application-evaluation guide. If the claim comes from a tool-using agent demonstration, the agent-demo checklist examines the surrounding tools, state and checked outcome. To go from written explanations to experiments on internal computations, read the mechanistic interpretability guide.
Mechanistic interpretability explained: five original reads
Twitter shadowban test: check search, replies and labels on X
From a vibe-coded prototype to a checked change