Adaptive reasoning: when to skip chain-of-thought

What adaptive reasoning is

Adaptive reasoning means choosing how much the model should reason before you send the task: no reasoning, a short capped chain, or a full step-by-step breakdown. It is the counterpart to chain-of-thought prompting: CoT explains how to make reasoning work, adaptive reasoning decides when to spend it.

The core observation is that a model has no built-in switch between recalling and reasoning; it reads the switch from the prompt. A study of "remember" versus "think" modes describes how words like "analyze", "you are an expert" or "consider all options" push a simple factual question into reasoning mode, where the model starts doubting the obvious; it reports that mixing up the two modes costs up to 40% of resources and lowers accuracy.

So the decision is a routing decision. Reasoning pays for itself when the task has a verifiable answer and several dependent steps. It tends to waste tokens on recall, and it actively hurts tasks where more text means more drift.

Why more reasoning can make answers worse

Three findings explain why less reasoning can mean better answers.

• No internal verifier. On math each step can be checked; on ethics, strategy or one-off forecasts it cannot. A preregistered study of the meta-uncertainty regime across 7 models and 79 open-ended tasks found that extended CoT cut accuracy by 17.3 percentage points, while control tasks with definite answers showed no such drop. Each extra step added another viewpoint rather than more signal.
• Drift during long chains. When picking from a fixed set of options, the model can talk itself from one option into another. Function-Routing CoT reports that a little reasoning gave +45% while long reasoning fell below the no-reasoning level, and fixes it by naming the choice on the first line.
• Literal instruction following. Reasoning models do what the wording asks. In a preregistered benchmark of coding agents, "develop several approaches and compare them" inflated reasoning tokens 2.4–7.4x and "think deeply" 1.6–2.2x, with no change in the final code.

Model strength shifts the line. Across 23 prompting techniques on 12 models, strong models did best with a clean, direct prompt, while budget models gained from short structuring moves such as Step-Back and Rephrase. In a GSM8K test, a strict rule-heavy prompt helped GPT-4o but cost GPT-5 2.4 points against simple CoT.

How to apply it

The practical toolkit has four parts.

1. Route by task type. Selective reasoning classifies each request first: facts, definitions and lists get a direct answer; math, cause-and-effect and trade-offs get CoT. On 12,000 MMLU-Pro questions this gave +10.2% accuracy, -47% latency and -48% tokens versus always or never reasoning. Select-then-Solve extends the map: direct requests for code and simple facts, plan-then-execute for multi-step research, draft-critique-revise for improving text.
2. Pick the mode yourself. Asking the model to choose its own strategy worked only for GPT-5 (67.1% vs 62.4% for a direct request) in the same study; the other models fell below baseline.
3. Cap or reorder when you keep reasoning. Post-reasoning puts the answer first and the justification after, which improved results in 88% of 117 model-task combinations. AdvPrompt uses evidence-style or role-based brevity instructions and reached 63-78% compression on simple problems without losing accuracy; threat-based prompts compressed only 4-25%. For coding agents, a scope, done-criterion and stop-condition template did not raise token use on any of the six models.
4. Prefer none over half. If the budget is tight, ask for a direct answer instead of "think briefly". In the broken-chains study, Gemini with CoT cut to 10% of its tokens scored 17%, against 67% with no reasoning at all.

Where to skip reasoning, and the limits

Where to switch reasoning off or keep it short:

• Summaries and digests. Ask for the result directly; step-by-step instructions led reasoning models to retell instead of compress, and raising the thinking budget of o1, o3 and GPT-5 in a summarization benchmark did not improve factual accuracy and sometimes lowered it.
• Simulating a person. For predicting how a specific persona reacts to content, intuitive prompting asks for an immediate reaction without cross-checking traits, because trait-by-trait logic slides toward stereotype.
• Ethical and strategic calls. Ask for a position in two or three steps and one deciding factor, as the meta-uncertainty study recommends.

Limits of the approach are real. The boundary between simple and complex is blurry for medium tasks, so the selective-reasoning authors suggest defaulting to reasoning when unsure. The meta-uncertainty effect was weaker on frontier models and disappeared on the smallest ones, which benefit from any reasoning. And most numbers come from math, code, QA or support benchmarks; test on your own tasks before fixing a rule.

You are a {role, e.g.: data analyst, developer assistant, marketer}.
Before answering a question:
1. Determine the task's complexity:
   - SIMPLE: a fact, a definition, a list, well-known information, a standard process
   - COMPLEX: math, logic, cause-and-effect, trade-offs, multi-step analysis, non-obvious patterns
2. Choose the response mode:
   - For SIMPLE → give a direct answer with no reasoning
   - For COMPLEX → use chain of thought:
     a) break the problem into parts
     b) analyze each part
     c) synthesize a conclusion
Question: {user_question}

{task context}
This is a {task type: ethical dilemma / strategic decision / one-of-a-kind forecast} — 
there is no single correct answer here.
Don't spread your reasoning wide. Answer in {2-3} steps max.
Pick out {one main factor / one position / one conclusion}.
Don't hedge, don't list "on the one hand — on the other."

Task: {short description of the problem}
Scope: change only {specific files/functions}. 
Do not touch any other code.
Done criterion: {what must work / which tests must pass}
Stop condition: as soon as the criterion is met, stop. 
Do not look for extra improvements and do not suggest alternative solutions.