From chain-of-thought to reasoning models: giving LLMs scratch paper
TL;DR
Chain-of-thought has models write intermediate steps and sharply improves accuracy on multi-step problems; reasoning models build this in. With current Claude, use thinking and effort rather than forcing reasoning into the prompt.
Give an LLM a multi-step word problem and force an immediate answer, and it often gets it wrong. Let it "think on paper" step by step first, and accuracy jumps. This seemingly simple finding, called chain-of-thought (CoT), led directly to today's reasoning models, which have their own "thinking" mode. This post covers three foundational studies and how to apply them correctly with modern models.
Three milestones
| Study | Idea | Main finding |
|---|---|---|
| Chain-of-Thought Prompting (Wei et al., 2022) | Give a few examples that spell out the reasoning steps | Big gains on maths, commonsense and symbolic reasoning; the benefit appears mainly in sufficiently large models |
| Zero-shot reasoners (Kojima et al., 2022) | Just add "Let's think step by step", no examples | Accuracy on MultiArith rose from 17.7% to 78.7% with the model tested |
| Self-consistency (Wang et al., 2022) | Sample several reasoning paths and take the most common answer | Clearly better than a single chain on arithmetic problems |
Why does "thinking on paper" work?
A model generates one token at a time, with a fixed amount of computation per token. Writing out intermediate steps is like giving it scratch paper: the result of each step becomes context for the next, splitting a big problem into manageable small steps.
From prompts to reasoning models
Newer models are trained to reason in a separate thinking block before answering. What that means for prompting:
- For current Claude models, Anthropic's docs advise relying on built-in thinking (with a suitable effort level) rather than asking the model to write its reasoning out in tags. On some of the newest models, asking it to write reasoning in tags such as
<thinking>may be declined. - Manual CoT is still a fallback when thinking is off: ask it to think step by step and put the final answer in
<answer>tags for easy extraction. - Self-consistency is still useful for important decisions: run several times and compare.
When is long reasoning worth it?
| Use it for | Not needed for |
|---|---|
| Multi-step maths, logic, planning | Simple fact lookups |
| Debugging, root-cause analysis | Classification, well-defined extraction |
| Comparing constrained options | Rewriting, translation, short summaries |
Risks and caveats
- Reasoning traces aren't faithful explanations: the steps a model writes may not reflect what actually produced the answer.
- Token and time cost: long reasoning raises cost and latency; adjust effort to difficulty.
- Overthinking: on simple tasks, long reasoning can talk the model into a wrong answer.
How to apply it
- Dev: Turn on thinking for reasoning tasks and keep effort low for simple ones; measure quality and cost.
- Tech Lead: For important automated decisions, run 3–5 times and act only when the runs agree.
- PM: Don't show raw reasoning to users as "proof"; show sources and verification results instead.
Sources & references
- Chain-of-Thought Prompting Elicits Reasoning in Large Language Models — Wei et al. (arXiv)
- Large Language Models are Zero-Shot Reasoners — Kojima et al. (arXiv)
- Self-Consistency Improves Chain of Thought Reasoning — Wang et al. (arXiv)
- Prompting best practices — Claude Docs