The Generational Bottleneck
Imagine playing a game of "Telephone," but instead of passing a single phrase, you are tasked with learning an entire artificial language from a speaker, and then teaching your version of that language to a new student. This recursive process—where the output of one learning agent becomes the primary training data for the next—is known as Iterated Learning.
Intuitively, one might expect information passed down such a chain to degrade into meaningless noise. However, cognitive science and machine learning research demonstrate the exact opposite: iterated learning acts as a powerful evolutionary filter. When individuals learn from imperfect data, their mistakes are not random. Instead, their reproductions are systematically distorted by their internal cognitive constraints, or inductive biases. Over multiple generations, these subtle individual preferences accumulate, stripping away idiosyncratic noise and transforming unstructured data into highly regular, learnable systems.
Iterated Learning as a Markov Chain
To understand why this convergence occurs, we can formalize the generational transmission of knowledge mathematically. As shown by Michael L. Kalish et al. in Iterated learning: Intergenerational knowledge transmission reveals inductive biases, iterated learning can be modeled as a Markov chain.
In each generation, an agent observes data $x$, infers a hidden hypothesis $h$ via a posterior distribution $P(h|x) \propto P(x|h)P(h)$ (where $P(h)$ represents the agent's prior or inductive bias), and then generates new data $x'$ from that hypothesis. Because the data at generation $t+1$ depends only on the data from generation $t$, the sequence of transmitted information forms a Markov chain over the hypothesis space.
Thomas L. Griffiths et al. proved in Connecting human and machine learning via probabilistic models of cognition that as this chain runs over many generations, the influence of the initial starting data decays to zero. The system inevitably converges to a stationary distribution that directly reflects the learners' underlying prior distributions.
This mathematical property allows researchers to use serial reproduction chains as an experimental tool to "sample" and map out latent human categories. For instance, Jing Xu et al. demonstrate in How memory biases affect information transmission: A rational analysis of serial reproduction that serial transmission acts like a Markov Chain Monte Carlo (MCMC) algorithm, revealing the internal probability distributions that define human concepts. Similarly, Stephan C. Meylan et al. leverage this mechanism in Evaluating models of robust word recognition with serial reproduction to isolate the abstract hierarchical structures humans use to decode noisy speech.
Solving the Poverty of the Stimulus
This mechanism fundamentally re-frames a classic debate in linguistics. For decades, the "Nativist" view argued that because children learn complex languages from sparse and messy input, they must possess a massive, innate "Universal Grammar." However, Willem H. Zuidema et al. challenge this in How the Poverty of the Stimulus Solves the Poverty of the Stimulus, showing that the transmission bottleneck itself solves the problem from the outside in. Because language must survive the filter of being learned by successive generations, the language itself evolves to become more structured and easier to learn.
When information is forced through this bottleneck, several structural properties spontaneously emerge:
- Regularization: Even if individual learners have only a weak preference for consistency, passing data through a chain rapidly eliminates variation and noise, as demonstrated by Florencia Reali in The evolution of frequency distributions: Relating regularization to inductive biases through iterated learning and explored further by the authors of The cognitive roots of regularization in language.
- Compositionality and Feature Economy: Holistic, unstructured mappings naturally break down into systematic, compositional systems where reusable sub-units represent specific features (Henry Brighton et al. in Language as an evolutionary system; Tessa Verhoef et al. in Cultural Emergence of Feature Economy in an Artificial Whistled Language). As Rahma Chaabouni et al. note in Compositionality and Generalization In Emergent Languages, while deep learning agents can communicate using messy, entangled codes, compositional languages are significantly more stable across generations because they are easier to transmit.
The Dual Pressures of Learning and Communication
If iterated learning purely optimizes for learnability and simplicity, why don't languages collapse into a single, effortless word for everything?
As Bill Thompson demonstrates in Human Biases Limit Cumulative Innovation, unconstrained cognitive biases can act as a restrictive leash that caps innovation and pulls systems toward oversimplified prototypes. To prevent this collapse and produce functional systems, a counter-pressure is required: interpersonal communication.
The emergence of optimal conceptual structures requires a delicate balance between two competing forces:
- Iterated Learning (Simplicity): Pushes the system to be compressible and easy for an individual to learn (Morten H. Christiansen et al. in Cultural Evolution of Language: Implications for Cognitive Science).
- Communication (Informativeness): Pushes the system to be precise and useful for transferring information between distinct individuals.
Jon W Carr et al. show in Conceptual structure is shaped by competing pressures for simplicity and informativeness and The Cultural Evolution of Structured Languages in an Open-Ended, Continuous World that true linguistic optimality—such as continuous, convex semantic categories—only emerges when both pressures interact. This dual-pressure dynamic explains why human color-naming systems are highly efficient, a finding replicated via neural network simulations by Emil Carlsson et al. in Iterated learning and communication jointly explain efficient color naming systems and E. Carlsson in Cultural evolution via iterated learning and communication explains efficient color naming systems. In population turnover models, such as the replacement microsocieties studied by Kenny Smith in Cultural Evolution and Perpetuation of Arbitrary Communicative Conventions in Experimental Microsocieties, this balance allows arbitrary conventions to become simpler and more efficient over time while remaining highly communicative for the group.
Engineering Systematicity in Artificial Intelligence
In modern artificial intelligence, models often struggle with systematic generalization, frequently memorizing idiosyncratic patterns in training data rather than learning compositionality. To combat this, computer scientists have adapted iterated learning into a training framework called Neural Iterated Learning (NIL).
By forcing successive generations of neural networks to learn from the communication data produced by their predecessors, researchers introduce an artificial transmission bottleneck. Shangmin Guo et al. demonstrate in Compositional languages emerge in a neural iterated learning model that this generational pipeline naturally selects for languages that generalize better to unseen concepts. Similarly, Ankit Vani et al. apply this in Iterated learning for emergent systematicity in VQA to nudge neural module networks into developing structured, compositional internal programs for visual question answering.
This process can even be simulated entirely in-context without updating model weights. Nathaniel Imel et al. show in Culturally transmitted color categories in LLMs reflect a learning bias toward efficient compression that treating the output of one Large Language Model generation as the few-shot context for the next creates an "iterated in-context learning" chain. This recursive prompting process reveals that LLMs possess an inherent inductive bias toward efficient compression, naturally restructuring random naming systems into near-optimal, human-aligned categories over time.
Ultimately, this framework opens up a "social path" to artificial general intelligence. As argued by Suzanne Sadedin et al. in A social path to human-like artificial intelligence, moving beyond static datasets toward multi-agent systems with generational turnover enables cumulative culture. By transmitting knowledge across successive generations of AI learners rather than relying solely on direct human experts (Avishkar Bhoopchand et al. in Learning few-shot imitation as cultural transmission), multi-agent populations can achieve distributed computation and rational information accumulation (Mathew D. Hardy et al. in Overcoming Individual Limitations Through Distributed Computation: Rational Information Accumulation in Multigenerational Populations), allowing complex cognitive traits and linguistic conventions to co-evolve dynamically (Marieke Woensdregt et al. in A model of cultural co-evolution of language and perspectivetaking).
Key Foundational Work
The evolution of vocabulary (2004) — contributed 2 concept elements to this direction.
How the Poverty of the Stimulus Solves the Poverty of the Stimulus (2002) — contributed 1 concept elements to this direction.