Reasoning
DeepMind framework ↗At a glance
Drawing valid conclusions: deduction, induction, analogy and maths.Reasoning is drawing valid conclusions by applying logic: deduction, induction, analogy and mathematics.
The Wason selection task shows how framing matters: people often fail the abstract version but solve the same logic framed as a social rule.
- Human ref.
- 69.7
- Best AI
- 96.4
- Gap
- -26.7
- Our items
- 15
Sub-faculties
The parts of this faculty, what our items cover today, and what is still to come
| Sub-faculty | Mirrors | Items | Status |
|---|---|---|---|
Deductive reasoning Drawing only the conclusions the rules allow, including which case could prove a rule wrong. How it differs: The rule is stated, so nothing has to be learned or guessed; the work is in following its logic strictly. In rule induction and in learning from feedback, the rule is hidden and has to be found. Compare all four → Mirrors Wason selection task | Wason selection task | 6 in the 5-min test | Covered |
Quantitative reasoning Reasoning with numbers and quantities, beyond plain arithmetic. Mirrors Cognitive reflection test | Cognitive reflection test | 6 in the 5-min test | Covered |
Analogies Seeing that one relationship matches another: A is to B as C is to what? Mirrors Webb, Holyoak & Lu (2023) | Webb, Holyoak & Lu (2023) | 3 | Covered |
Human vs AI
Best published result on each side, 0–100
Humans. Experts do well on hard questions in their own field. On the abstract Wason task, most people pick the card that confirms the rule instead of the one that could break it.
AI. Frontier models now outscore PhD experts on graduate-level science questions. Textbook logic puzzles are widely published online, so scores on them say little.
- Human ref. 69.7
- Share of GPQA Diamond's graduate-level science questions answered correctly by PhD holders in the question's own field, as reported by OpenAI. About the test →
- Best AI 96.4
- The same questions answered by Gemini 4 Argon at its highest setting, in Epoch AI's own run rather than the developer's (October 2026). About the test →
Example items
Specimens written for this page. Live items are similar but never published.
Pick the word that completes the analogy.
(a) Question
- 1. Sculptor is to statue as baker is to …(a) oven(b) bread(c) flour(d) kitchen
Read the problem, then type the number.
(a) Stimulus
(b) Question
How many cents does the ball cost?
Which cards must you turn over to check whether the rule is broken?
(a) Rule
(b) Cards
Benchmarks
| Benchmark | Origin | Items | Best AI | Humans |
|---|---|---|---|---|
| Webb, Holyoak & Lu (2023) | Webb, Holyoak & Lu (2023), Nature Human Behaviour, 2023 | 3 | 3 of 4 | — |
| Cognitive reflection test | Cognitive reflection test (Frederick, 2005), 2005 | 6 | 96% | 38% |
| Wason selection task | Wason selection task, 1968 | 6 | n/a | < 10% |
In everyday life
- Checking whether a rule was broken
- Working out a fair split of a bill
- Spotting a flawed argument