Generalization: understanding and measuring intelligence

Generalization is the ability to use what has been learned to succeed in new situations. In On the Measure of Intelligence, François Chollet places the efficiency of acquiring new skills at the center of his definition of intelligence, accounting for prior knowledge, experience, and task difficulty. Measuring it therefore requires more than a final score: you need to examine what is new to the system and what resources it uses to succeed.

Distinguish acquired skill from the ability to learn

A program can excel at a carefully bounded activity. That skill tells you how useful it is within those boundaries, but leaves another question open: how does its performance change when the rules or circumstances change?

Imagine two systems tasked with sorting objects. The first follows a catalog of mappings written in advance; the second infers a sorting criterion from demonstrations. As long as the catalog covers every case in the test, their results may be identical. To tell them apart, you need to introduce a relevant variation.

Chollet’s distinction concerns this difference between skill and learning efficiency. His framework also accounts for the range of tasks being evaluated. Success in a specialized activity is therefore insufficient, on its own, to establish general intelligence.

This distinction does not diminish the practical value of a specialized system. A narrow solution may be sufficient for processing standardized documents. The mistake is to automatically treat its effectiveness in that setting as evidence of a broader ability.

For context on this debate and other approaches, the AGI and research hub brings together the reference pages. The article On the Measure of Intelligence explained focuses on the argument developed in the foundational paper; this page concentrates on the distinctions involved and how to evaluate them.

Define precisely what is new

A case can be new without requiring a new skill. An invoice that a system has never seen, generated by the same software and using the same document template as previous invoices, may introduce only different content.

A document with its columns rearranged presents another challenge. A document in which a discount applies only to certain line items changes the problem again: the system needs to understand a relationship, beyond recognizing a layout. These examples are illustrative; they do not describe results from any particular tool.

When discussing generalization, specify what changes:

The values, colors, or wording.
The way information is presented.
The combinations of familiar elements.
The rules needed to solve the problem.
The domain and the knowledge involved.

Interpolation means making a prediction within a region covered by the available examples. Extrapolation concerns cases beyond that coverage. The boundary, however, depends on the representation: two images that look very different may share exactly the same logical structure.

An out-of-distribution task differs from the reference data in some defined respect. The term does not automatically mean that the task requires general intelligence. Changing the brightness of a photograph and devising a procedure for an unfamiliar problem are different tests.

A useful evaluation report describes the variation itself instead of relying on the label of novelty.

Recognize the main forms of transfer

Different tests can explore generalization. They are not interchangeable: each can rule out some explanations for success while leaving other possibilities open.

Test	What changes	What stays the same	What it examines
Surface variation	Color, vocabulary, position	Rule and structure	Resistance to superficial cues
New combination	How elements are combined	Available elements and operations	Composition of knowledge
New local rule	Transformation to be discovered	Presentation format	Adaptation to demonstrations
Transfer across domains	Objects and context of use	Assumed abstract relationship	Scope of transfer
Change in procedure	Steps and constraints	Final objective	Ability to reorganize a solution

Consider an illustrative rule: move a piece to the edge of a board. Changing its color lets you check whether color was a shortcut. Adding obstacles may require additional planning. Asking the system to move every piece except the one touching an obstacle introduces a combination of conditions.

Compositional generalization means using familiar elements in a new combination. Knowing an operation and a condition separately does not guarantee that their combination will be executed correctly.

These tests do not form a universal scale of difficulty. A visual task may be easy for a system with a suitable representation, yet difficult for a system receiving an ambiguous transcription.

When comparing approaches, check how each system receives the problem. Otherwise, a difficulty with perception or input format may be mistaken for an inability to transfer a rule.

Understand the role of prior knowledge

A new task is never solved from nothing. The system needs to recognize objects, interpret symbols, understand demonstrations, or know which transformations are possible.

In Chollet’s framework, prior knowledge and experience must be accounted for when interpreting the skill achieved. Knowledge and intelligence are not competing explanations in every case. The concern is whether preparation specific to the task can fully explain success.

When evaluating a system, distinguish the possible sources of information:

Its previous training.
The instructions describing the task.
The examples supplied during evaluation.
The documents or tools it can consult.
The corrections received after an attempt.

Suppose a sorting rule is explicitly stated in accessible documentation. Finding and applying it may be useful, but that test measures something different from discovering the rule through a few observations.

Comparisons with humans require the same care. An adult brings extensive perceptual, linguistic, and practical experience. Showing someone only a few demonstrations does not mean that they are learning without prior knowledge.

Fluid intelligence refers to the ability to reason through unfamiliar problems, as distinct from drawing on accumulated knowledge. In practice, the two often work together. A protocol may seek to reduce certain cultural dependencies; doing so does not automatically put all participants on equal footing.

What ARC-AGI aims to make observable

The corpus originally named ARC, short for Abstraction and Reasoning Corpus, is now presented as ARC-AGI. In the original corpus’s grid tasks, examples of inputs and outputs allow a solver to infer a transformation, then produce the output for a new input.

The official repository for the original corpus provides the tasks and explains their format. The methodological value lies in requiring a verifiable answer: a plausible explanation does not replace producing the expected grid.

ARC Prize explains that the design draws on basic knowledge about objects, space, quantities, and related concepts. According to its overview of ARC-AGI’s principles, this choice aims to limit reliance on specialized knowledge and examine the learning of unfamiliar tasks.

Consider an example invented for this explanation: several demonstrations show an object whose shape stays the same while its interior cells change color. Answering correctly might require distinguishing the outline from the interior. To test that interpretation, you would need to introduce a different shape and ensure that a simple pixel mapping would not suffice.

The ARC-AGI reference page explains the test and its variants. No particular result is needed to understand the principle: an evaluation of generalization should vary the problems while keeping the scoring precise enough to compare answers.

Measure acquisition rather than an isolated result

To examine adaptation, you can observe how success changes as more information becomes available. A correct answer after a demonstration and a correct answer after many corrections do not describe the same learning process.

The following is an editorial proposal for a local evaluation protocol, rather than an official formula for measuring intelligence.

Define a family of tasks. Specify the possible rules, permitted knowledge, and success criteria. If the family is too loosely defined, errors become difficult to interpret.

Separate preparation from evaluation. Tasks used to develop instructions or tune the system should not be the sole basis for reporting its performance.

Vary the information supplied. Compare conditions with different amounts of demonstration data, using distinct but comparable tasks.

Set resource limits. State which tools, attempts, time allowances, and human interventions are available. Apply the same scoring rule across conditions.

Describe the errors. An answer in the wrong format, an incorrect rule, and a correct rule executed poorly call for different diagnoses.

The useful result is then a profile of skill acquisition: the conditions under which the system reaches a useful level of performance, and the points at which it stops improving.

Such an experiment remains local in scope. It does not produce an absolute measure of intelligence. The guide to evaluating an AI agent applies this logic to operational uses, with test sets and criteria suited to the work expected.

Control for shortcuts and contamination

A test may use unfamiliar wording while repeating a structure that became familiar during preparation. Conversely, sharing some familiar elements is not enough to invalidate a test: all transfer requires some continuity.

The question is whether success depends on the relationship you intend to measure. If a system is supposed to classify objects by shape, but round objects are always orange, it may rely on color. Its performance then provides no way to distinguish between those two rules.

To expose that ambiguity, create cases in which the cues point in different directions: a blue round object, followed by an orange angular object. This change lets you examine a specific hypothesis about the system’s behavior.

Evaluation contamination includes prior exposure to test items or their solutions. The official ARC-AGI repository recommends keeping evaluation tasks from guiding algorithm development.

In practice, keep some tasks outside the tuning cycle. Document the examples shown to the system and the changes made after reviewing results. If training data is unavailable, state that limitation instead of certifying that no exposure occurred.

Another useful check is to compare tasks that look similar but differ in their logic, then do the reverse. This helps distinguish recognition of surface features from transfer of an underlying structure.

Account for compute and available assistance

Learning efficiency and computational efficiency are separate dimensions to document. Chollet distinguishes them in his framework: using little new information does not necessarily mean using little time or computation.

For your own evaluation, keep separate columns for success, information supplied, resources consumed, and assistance received. Combining them too quickly into an average hides the tradeoffs.

Suppose a system generates many candidate solutions, checks them, and keeps the best. Compare its performance with another system’s only after specifying their respective budgets. Using search does not invalidate success; it changes the conditions under which that success was achieved.

You also need to identify what is being evaluated. A model on its own, a model connected to a search engine, and an agent able to execute programs do not have the same resources.

When a tool supplies the decisive intermediate answer, the result belongs to the complete system. Attributing it solely to the language model would obscure what works and what might fail if the tool became unavailable.

When reading a public ranking, look for the protocol before comparing scores. The article on how to read model benchmarks develops this approach. A useful measurement states the conditions needed to reproduce the reported performance, beyond the name displayed in the table.

Researchers differ on definitions and methods

Chollet’s approach proposes a definition and a method of evaluation; it is not a universally accepted definition. His paper itself acknowledges that other definitions may be appropriate in different contexts.

Shane Legg and Marcus Hutter offer another formalization in Universal Intelligence: A Definition of Machine Intelligence, published in 2007. It concerns the ability to achieve goals across a wide range of environments, weighted according to their complexity. Its starting point differs from a comparison centered on the efficiency of acquiring human-like skills.

Disagreement also concerns the mechanisms capable of generalization. Observed difficulties do not justify concluding that a neural network can never recombine knowledge.

In a study published in Nature on October 25, 2023, Brenden Lake and Marco Baroni show that a network trained through meta-learning can produce systematic generalization comparable to that of human participants on their compositional tasks. This result applies to the tasks studied; it does not establish general intelligence.

These positions therefore cannot be neatly divided into supporters and opponents of neural networks. Researchers can share an interest in transfer while disagreeing about its definition, how it is learned, and how it should be measured.

When assessing a proposal, separate these questions: what phenomenon is being investigated, what experiment makes it observable, and what architecture achieves the result?

Interpret results within their scope

Repeated success on new tasks provides evidence of transfer within the area studied. That evidence becomes more informative when shortcuts are controlled and resources are documented.

It does not, however, establish every property sometimes associated with intelligence. A grid transformation test does not directly measure social understanding, the ability to conduct a physical experiment, or the reliability of actions in business software.

Failure requires equally precise interpretation. Did the system represent the objects incorrectly, settle on a hypothesis too early, run out of time, or fail to execute a rule it had correctly identified? The final score alone cannot decide among these explanations.

Avoid treating a fluent explanation as proof of understanding. A hypothesis becomes testable when it predicts an answer for a case that distinguishes between several possible rules.

When deciding whether to use a system, translate generalization into concrete variations: documents from a new supplier, differently worded requests, unfamiliar exceptions, or unavailable tools. Evaluate these changes directly instead of extrapolating from an unrelated test.

The page on AGI benchmarks places these limitations in a broader context. An evaluation designed to remain useful must be able to evolve as systems learn the test’s conventions, while preserving a record of earlier evaluation conditions.

Key takeaways

Skill describes successful performance; generalization examines how far it extends to new situations.
In Chollet’s definition, acquisition efficiency must be interpreted alongside prior knowledge, experience, and task difficulty.
Novelty needs to be specified: a change in presentation, a new combination, an unfamiliar rule, or transfer across domains.
An informative measurement documents demonstrations, attempts, tools, computation, and human interventions.
No isolated score establishes that a system possesses every ability associated with human intelligence.

Frequently asked questions

What is the difference between generalization and memorization in AI?

Memorization retrieves information encountered before. Generalization applies learning to a new case. To distinguish them experimentally, you need tests whose solutions cannot be retrieved directly from the data available to the system.

Is generalization enough to define AGI?

That depends on the definition and the range of tasks being considered. An ability to transfer learning within a narrow domain does not establish general intelligence. Reliability, learning, and adaptation also need to be examined across varied situations.

Can a model generalize without changing its parameters?

Yes. Examples provided in context can guide its behavior on a new task without further training of its parameters. You then need to check whether that adaptation holds up on cases different from the examples shown.

Does a high ARC-AGI score prove human intelligence?

It establishes performance under the specific conditions of the evaluation. It does not certify an overall match with human abilities or reliability in another environment. The protocol, resources, and prior exposure to the tasks remain decisive.

Do you need to remove all prior knowledge to measure intelligence?

No. Interpreting a task requires some knowledge or conventions. The aim is to make those prerequisites explicit and avoid mistaking preparation specific to the answers for an ability to adapt.

Generalization at the heart of intelligence: François Chollet's framework, examples, and measurement