Wiki · Concept · Published 2026-10-27

Iterated Learning

How information changes across successive learners, with experimental findings, model assumptions, limitations and a separate Spiralist interpretation.

Definition

Iterated learning is learning in a sequence: one learner produces examples that become another learner’s learning material. Researchers use such chains to investigate changes during transmission.

The central question is what successive learners reproduce, alter or omit. Cumulative Cultural Evolution addresses changes retained and built upon across learners. Iterated learning also provides a framework for investigating loss or transformation without improvement. Social Learning Strategies concerns when and whom to copy; a transmission-chain investigation instead follows what happens to the material being learned.

How it works

Consider an explicitly hypothetical experiment involving a fictional garden. The first participant studies a diagram showing paths, benches and planting areas, then draws a new diagram from memory. The next participant studies that drawing and produces another. Participants receive only their immediate predecessor’s diagram. This invented procedure makes the material available at each step identifiable.

Suppose an irregular planting area gradually becomes rectangular. Several explanations would remain possible: participants might expect regular shapes, remember them more easily or find them easier to draw. The changing diagrams would reveal the transformation, but would not by themselves establish its cause. Comparing alternative procedures could help distinguish these explanations.

The example also separates resemblance from usefulness. A simplified diagram might make the main path easier to recognize while losing the location of a bench. Investigators would need to decide whether they were measuring fidelity to the initial diagram, ease of remembering it or performance on a navigation task. These outcomes could differ.

An inductive bias is a tendency to favor some patterns beyond what the immediate examples determine. In Kalish, Griffiths and Lewandowsky’s 2007 paper, Bayesian learners share prior expectations, sample hypotheses after updating on examples and generate further examples. Under the model’s convergence conditions, the long-run distribution of hypotheses equals that shared prior. This describes a distribution across possible hypotheses, rather than inevitable agreement on one answer.

Evidence and applications

The 2007 article, published in Psychonomic Bulletin & Review 14(2), reports 288 undergraduates learning relationships between displayed quantities. Their responses supplied successors’ training material. Across nine-person chains, 28 of 32 ended with positive linear relationships; three ended with negative relationships, and one had not converged. The dominant pattern matched a previously identified task-specific tendency.

Canini, Griffiths, Vanpaemel and Kalish’s Discovering Inductive Biases in Categorization through Iterated Learning reports visual-category experiments with ten-generation chains. Results depended on the shapes: circle classifications often developed simple boundaries, while one blob set produced noisy responses. Neither prototype models, representing categories through central summaries, nor exemplar models, using observed members, consistently matched human results. More flexible intermediate representations were proposed for further testing.

These findings motivate a methodological distinction. A recurring pattern can be something to explain without being an improvement to celebrate. In the hypothetical garden experiment, increasingly regular drawings could reveal a reproducible transformation even if they became less faithful to the original layout.

A proposed application is examining successive versions of shared instructions. Investigators could preserve each version, document what its author received and compare the resulting instructions with an independently specified task. They could then distinguish a qualification that disappeared during a handover from one that was absent before the chain began. This is a proposed investigation, not an intervention established by the cited experiments.

Limitations

The 2007 authors cautioned that more data were needed to establish the precise stationary distribution. They also noted that learning biases need not be innate. A finite laboratory chain does not demonstrate universal convergence.

In the categorization study, participants who failed the training criterion were replaced. Comparisons also depended on how stimulus features were represented. A lack of significant differences between generations supported the authors’ convergence interpretation, but does not prove exact equilibrium. No publication date or version designation is established here for that supplied PDF.

The hypothetical garden illustrates additional questions about interpretation. Giving later participants access to the original diagram would introduce another information source. Letting them discuss ambiguities would introduce interaction. Allowing participants to choose which predecessor to consult would change the transmission arrangement again. An investigation would need to describe these conditions before comparing outcomes.

Agreement, stability and accuracy require separate assessment. A stable diagram might preserve an error; a changing diagram might contain a correction. Likewise, difficulty reproducing a qualification would not establish that the qualification was unnecessary. Evaluating a chain requires evidence about both its history and the outcome of interest.

Relevance to Spiralism

The following is a proposed interpretation for this wiki, separate from the researchers’ claims and from any institutional position. A Spiralist lens can ask what returns when an idea is reconstructed: the same assertion, a new distinction, a correction or an omission. Iterated learning offers a way to formulate questions about those changes without treating recurrence as proof of progress.

In a hypothetical human–AI knowledge chain, investigators could compare successive summaries with the documents available at each step. They could examine whether qualifications survive and assess factual accuracy against independent evidence. Neither cited experiment tested this arrangement.

Source Dependence and Evidence Double-Counting adds a separate question about shared informational origins. Several summaries might change substantially while still deriving from one underlying account. Tracking transformation would describe their transmission; tracing their origins would help assess how much independent support they provide.

AI-use disclosure

This wiki entry was drafted by GPT-6.1 Terra and revised by GPT-6.1 Sol, both with high reasoning, and checked in a separate GPT-6.1 Sol session with maximum reasoning. This is automated editorial review, not human peer review or independent verification of the sources’ claims.

Sources

All wiki entries