Underthinking is premature exit from internal reasoning on hard multi-step problems — the model answers confidently while budget and task difficulty still demand more deliberate search.
Underthinking is like handing in a math test early because you felt confident, even though three multi-step problems needed more scratch work. The model stops its internal thinking trace too soon on a genuinely hard question, while overthinking is the opposite — wasting pages on an easy one. Billing and GRPO details are unrelated distractors.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Reasoning models introduced two symmetric failure modes on the thinking budget dial: overthinking and underthinking. This MCQ tests the latter and filters distractors about billing and GRPO architecture.
Underthinking is a policy behavior — not an API pricing artifact.
Defining underthinking
Underthinking occurs when a reasoning model stops internal deliberation too soon on a problem that still requires extended multi-step search — and does so despite remaining thinking budget.
Symptoms: short hidden traces on hard items, confident wrong final answers, skipped verification steps, failure on problems needing backtracking.
The model had room to think longer but the decoding policy terminated early — mis-calibrated difficulty estimation or reward hacking toward short answers.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- OpenAI o-series users report hard AIME items failed with surprisingly short reasoning traces — classic underthinking symptoms.
- DeepSeek-R1 eval discussions note early-stop behaviors on multi-step proofs when thinking budgets are low.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does RLVR training interact with underthinking?
Outcome-only rewards may encourage short lucky paths — process rewards or length-aware shaping can penalize premature answer emission on hard items.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Picking option B — that describes overthinking (too many tokens on easy tasks), the mirror failure, not underthinking.
60 second bullets to scan on the way to the call.
Define underthinking in one sentence
Contrast with overthinking
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.