nucleation

LIT-820: Iterated learning: Intergenerational knowledge transmission reveals inductive biases

3 min read

Properties4
StatusActive — read 2026-10-09 (NOTE-616), in full, from the publisher's PDF; worth reading as the first laboratory test of iterated learning against an inductive bias known before the experiment: in 32 chains of nine human learners each passing a function-learning task down the line, 28 converged within a few generations to a positive linear function whatever the first learner was trained on (positive linear, negative linear, U-shaped or random pairings), as Bayesian agents sampling from a prior that favours positive linear functions would. The prior was not measured in the experiment but taken from earlier function-learning studies; 3 families ended on the negative linear function and 1 did not converge, and the authors say more data would be needed to show the chains had reached a stationary distribution. Bears on THEORY-155.
Tagscognition, probabilistic-modeling, linguistics
Cited as “paper” byNOTE-616 — Iterated learning (Kalish, Griffiths & Lewandowsky)
Cited as “source” byTHEORY-489 — Chains of human function learners, each trained on the previous learner's responses, forget their starting function and in 28 of 32 families converge within a few generations to a positive linear function, the bias earlier studies report, whatever function trained the first learner; this is convergence to the learners' favoured function, not shown to be convergence to their prior

Graph View

Michael L. Kalish, Thomas L. Griffiths and Stephan Lewandowsky (2007), Psychonomic Bulletin & Review 14(2):288–294 — DOI-10.3758/BF03194066

Key takeaways

  • The prediction. If learners share a prior p(h) and each samples a hypothesis from the posterior given the previous learner’s data, the chain of hypotheses is a Markov chain whose stationary distribution is the prior, so “the stimuli provided for learning are completely irrelevant in the long run” (citing Griffiths and Kalish, LIT-769). A simulation with Bayesian linear regression (appendix) shows the convergence in a few generations.
  • The test. 288 undergraduates in four conditions; each condition formed eight “families” of nine generations. The first learner in each family was trained on 50 points from y = x, y = 101 − x, a U-shaped sine function or a random one-to-one pairing; each later learner was trained on the previous learner’s 50 test responses (25 old and 25 new x values), without contact and without knowing their responses would be passed on.
  • The result. 28 of 32 families converged within a few generations to a positive linear function. All eight negative-linear families, and three others, produced a negative linear function for at least one generation; three families ended on it and one did not converge. The median correlation with y = x rose across generations in every condition except the positive linear one, which was at ceiling from the start.
  • Where the prior came from. The bias toward positive linear functions was taken from earlier function-learning studies (initial responses, ease of learning, a fitted model), not measured on these participants. The authors call the stationary distribution “apparently complex” and say significantly more data would be needed to confirm convergence and map it against prior estimates.
  • The claims drawn. Iterated learning can be used as a method to reveal implicit inductive biases; and the result is said to “validate” serial-reproduction studies such as Bartlett’s (LIT-768) as revealing biases, and to suggest that “languages, legends, religious concepts, and social norms” are tailored to them. Neither extrapolation is tested.

Standing in the record

Filed on 2026-10-09 from the manuscript bibliography of 2026-10-09 (work what-survives-translation) at the owner’s request: the manuscript considered it and dropped it from the final reference list. It is read here on its own merits. See the curation entry of that day.

Read on 2026-10-09 (NOTE-616). It is the empirical companion of Griffiths and Kalish’s analysis (LIT-769) and the closest thing the record holds to a test of THEORY-155: a human chain run against a bias stated before the experiment. It does not meet that theory’s promote_when, since the prior was not measured or set independently on the learners and stationarity was not shown. It stands beside Kirby, Cornish and Smith’s chains (LIT-771), which measured no prior at all.

Backlinks