AGI Index dimension · weight 12%

Generalization

Transfer of competence to genuinely unfamiliar problems.

Whether capability learned in one domain transfers to a domain the system was not trained for, without task-specific scaffolding. The dimension most directly implicated by the word "general" in AGI, and the hardest to measure because contamination is difficult to rule out.

The argument

Evidence and counter-evidence

Both sides are published at equal weight. A framework that only records what raises a score is not measuring anything.

Raises the score

Cross-domain transfer is observable: capability acquired in code appears to improve structured reasoning in unrelated symbolic domains.
Few-shot adaptation to novel task formats has improved substantially without fine-tuning.

Holds it down

Training-set contamination cannot be excluded for most public evaluations, so an unknown share of apparent transfer is recall.
Performance on deliberately novel abstraction tasks remains far below human baseline despite enormous gains elsewhere.
Transfer is asymmetric — it flows readily between text-adjacent domains and poorly into anything requiring physical intuition.

What the score reads from

Abstraction and reasoning corpora
Improving, still well short of unremarkable human performance.
Held-out task formats
Strong within modality, weak across.

Related coverage

Everything bearing on generalization

News

Research

Argument

Context

Where this sits against the rest

Full Index

How these scores are producedMethodology v1.2