Executive functions

DeepMind framework ↗
Human 92.0 · AI 74.5Humans lead +17.53 benchmarks →

At a glance

Planning, switching strategy when rules change, inhibiting old habits.

Executive functions steer thinking: planning, switching strategy when the rules change, and holding back habits that no longer fit.

The classic probe is card sorting: a hidden rule changes without warning, and the question is how fast you abandon what used to work.

Human ref.
92.0
Best AI
74.5
Gap
+17.5
Our items
10

Sub-faculties

The parts of this faculty, what our items cover today, and what is still to come

Sub-facultyItemsStatus
Rule switching
Dropping a rule that stopped working and finding the new one.
Mirrors Turn Bench
4Covered
Inhibition
Stopping an automatic response, like reading a word instead of naming its ink colour.
3Covered
Planning
Ordering steps so they reach a goal without dead ends.
Mirrors PlanBench
3Covered

Human vs AI

Best published result on each side, 0–100

Human ref.
92.0
Best AI
74.5

Humans. University-educated testers complete most of GAIA's multi-step assistant tasks.

AI. The best agents complete fewer of GAIA's tasks than people do. When a game's rules flip in secret, some models notice but keep reconciling the change with their old picture of the game.

Human ref. 92.0
Share of GAIA's multi-step assistant questions answered correctly by university-educated testers, in the GAIA paper (2023). About the test →
Best AI 74.5
Share answered correctly by the best agent on Princeton's HAL leaderboard, Claude Sonnet 4.5, where every agent runs under the same harness (May 2026). About the test →
255075100Perception+13Generation-35Attention+44Learning0Memory+21Reasoning-27Metacognition-55Executivefunctions+17Problemsolving+2Socialcognition+4
Human referenceOur estimateBest published AILabel = human − AI, points

Example items

Specimens written for this page. Live items are similar but never published.

Catch the rule changespecimen · not scored

Sort each card onto a pile. Only “right” or “wrong” is shown; the hidden rule changes without warning.

(a) Piles

(b) Card to sort

Feedback: “right” / “wrong” only

Figure 1. A rule-switching item. The sorting rule (colour, shape, number) switches after four correct sorts. Benchmark: Turn Bench. Answer: Two red squares: pile 1 under the colour rule, pile 3 after the switch to shape, pile 2 once number rules.
Turn Bench benchmark page →
Restack the blocksspecimen · not scored

Move one top block at a time until the towers match the goal.

(a) Start

AB
C

(b) Goal (3 moves at least)

BAC
Figure 2. A block-stacking plan. C has to end on top, so it has to move last. Benchmark: PlanBench. Answer: 3 moves: B onto the empty tower, then A on top of it, then C.
PlanBench benchmark page →
Name the inkspecimen · not scored

Name the colour of the ink, not the word.

(a) One word at a time: name the ink

redgreenyellowblue
Figure 3. Stroop colour words. Reading is automatic, so a word that names a different colour slows people down and draws errors. Benchmark: Luo et al. (2025). Answer: Blue, green, red, yellow: the inks, from left to right.
Luo et al. (2025) benchmark page →

Benchmarks

3 rows
BenchmarkOriginItemsBest AIHumans
Turn BenchKaggle × Google DeepMind hackathon, Executive Functions track winner, 2026480%—
PlanBenchPlanBench (Valmeekam et al.), 2023397.8%—
Luo et al. (2025)Luo et al. (2025); Stroop (1935), 2025393.1%—

In everyday life

  • Adapting a plan when the train is cancelled
  • Stopping a habit that no longer helps
  • Juggling several deadlines