ModelRefs / Commonsense Reasoning — AI Glossary

Commonsense Reasoning — AI Glossary

The ability of a model to apply everyday world knowledge and physical/social intuitions that are rarely stated explicitly in text.

Overview

Commonsense reasoning requires knowledge humans acquire implicitly: objects fall when dropped, people feel sad at funerals, water is wet. Benchmarks: HellaSwag, WinoGrande, CommonsenseQA, PIQA, SocialIQA. LLMs trained on web text absorb much commonsense knowledge but still fail on out-of-distribution physical and social scenarios.

Reference details

Topic	evaluation
Last reviewed	2026-06-24

How it works

The knowledge at issue is the kind rarely written down because everyone has it: objects fall, containers hold things, people act on beliefs. Text corpora under-represent it precisely because stating it is unnecessary.
Models acquire much of it indirectly, from the implications of what is written rather than from assertions, which is why coverage is broad and uneven in ways hard to predict.
Evaluation works by constructing items where the commonsense inference is the only thing separating a right answer from a plausible wrong one — a sentence completion, a pronoun whose referent depends on world knowledge, a physical plausibility choice.
Adversarial filtering is how the wrong options are made hard: candidate distractors are generated and then kept only if a model finds them plausible, which removes items solvable by surface cues.
Nearly all of these benchmarks are multiple choice, which makes grading objective and installs a chance floor that varies by format — a binary item floors at 50%, a four-way item at 25%.
So scores are only comparable once corrected for that floor, and the raw percentages across a suite of commonsense benchmarks are not on the same scale.

Example: Why the same raw score means different things across these benchmarks

Two results, both 70%: one on a binary pronoun-resolution task, one on a four-way sentence-completion task. Correct each for chance by taking the distance above the floor as a fraction of the distance available. Binary: (0.70 - 0.50) / (1 - 0.50) = 0.20 / 0.50 = 0.40. Four-way: (0.70 - 0.25) / (1 - 0.25) = 0.45 / 0.75 = 0.60. The same raw figure represents 40% of the available headroom in one case and 60% in the other — the four-way result is half again as strong, and a table listing both as 70% says the opposite of that. The effect is largest near the floor: a raw 55% on the binary task corrects to (0.55 - 0.50) / 0.50 = 0.10, barely above guessing, while a raw 55% on the four-way task corrects to (0.55 - 0.25) / 0.75 = 0.40, four times as far above its floor. Averaging raw scores across a mixed-format suite therefore weights the binary tasks' noise most heavily, which is why a commonsense 'average' is close to meaningless unless every member shares a format.

Commonly confused with

Commonsense reasoning is not world knowledge and not logical reasoning, though it borders both. World knowledge is retrievable facts — capitals, dates, formulae — which a model either stores or looks up. Logical reasoning is inference from stated premises, where the premises are given. Commonsense is the unstated middle: the premises nobody supplies because they are assumed, which is why retrieval does not help and why a model can be strong at formal logic and still fail on whether a person can carry a piano upstairs alone.

When to use it

Reach for it when:

Assessing fitness for open-domain dialogue or assistance, where unstated assumptions are constant and users never supply them.
Checking robustness on physical or social scenarios relevant to an embodied or agentic application.
Diagnosing failures that look arbitrary, where the missing step is usually an assumption the model did not make.
Chance-corrected and within one format, so the numbers are on a single scale.

Reach for something else when:

Averaged raw across mixed formats, where binary and four-way floors are silently combined.
As a general capability measure, since these suites are largely saturated at the frontier and the remaining differences are small.
On public sets without considering contamination, which is the default assumption for widely-used benchmarks.
As evidence about a specific domain. General commonsense does not transfer to clinical, legal or industrial settings where the assumed knowledge is specialist.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Commonsense Reasoning — AI Glossary.

Frequently asked questions

What is Commonsense Reasoning?