Iterated Learning Model (ILM)

• Iterated Learning Model (ILM) is a computational framework that examines how structured languages emerge through constrained, intergenerational transmission.
• ILM employs methodologies ranging from Bayesian inference to neural autoencoding and policy gradient models to quantify expressivity, compositionality, and stability in communication systems.
• Empirical findings show that intermediate transmission bottlenecks foster stable, compositional codes, linking evolutionary language dynamics with principles from information theory.

The Iterated Learning Model (ILM) is a formal computational framework for studying the emergence and evolution of structure in communication systems—primarily human language—via intergenerational transmission constrained by cognitive and social bottlenecks. In standard ILMs, agents (e.g., neural, Bayesian, or modular networks) are arranged in a generational chain, each learning from the limited outputs of its predecessor. Over successive iterations, this mechanism reliably gives rise to expressive, compositional, and stable languages. Modern ILMs extend core principles to high-dimensional input spaces, LLM evolution, and multi-agent and population-level settings, and have established deep links to the Information Bottleneck principle, rate-distortion theory, and bias amplification dynamics.

1. Core Principles and Theoretical Foundation

The ILM formalizes the process by which a language is learned and transmitted across generations, with the central insight that a transmission bottleneck—learners only see a small set of meaning-signal pairs—exerts a strong inductive bias favoring compositional and expressive languages. Each generation's learner typically starts from a naïve or weakly biased state and must generalize from sparse exposure to reconstruct the mapping from meanings $M$ to signals $S$.

In Bayesian ILMs, the underlying framework is:

$P(h \mid d) \propto P_0(h) \, P(d \mid h)$

where $h$ is a hypothesis about the communication system (e.g., a grammar or mapping), $d$ is observed data, and $P_0(h)$ encodes the agent's inductive bias. After observing data sampled from the previous generation, the learner forms a posterior distribution, then samples a new hypothesis $h^*$, generating fresh data for the next generation. Over many generations, priors in $P_0(h)$ become exponentially amplified, causing the dynamics to converge on classes of languages determined by the structure of the bottleneck and prior (Ren et al., 2024).

Fundamental quantities of interest in ILM analysis include:

• Expressivity $x$, the fraction of meanings that receive unique signals,
• Compositionality $c$, degree of systematic mapping from meaning components to signal components,
• Stability $S$0, the fraction of signal-meaning mappings preserved between generations.

The transition bottleneck is the key parameter: extremely narrow bottlenecks yield unstable or degenerate systems; full exposure yields holistic, uninterpretable mappings; intermediate bottlenecks induce compositional, expressive codes (Bunyan et al., 2024, Sains et al., 2023).

2. Algorithmic and Architectural Realizations

Early work on ILM considered symbolic models with explicit mapping tables and "obverter" mechanisms for constructing encoders from learned decoders. More recent approaches operationalize ILM in neural architectures:

• Feedforward and Autoencoder Networks: Both encoding ($S$1) and decoding ($S$2) are parametrized as single- or multi-layer neural networks. The inclusion of unsupervised autoencoding phases—learning to reconstruct random meanings through the communication channel—prevents code collapse and supports expressivity and compositionality at scale (Bunyan et al., 2024, Lee et al., 6 Jan 2026).
• Policy Gradient and Modular Networks: For agents involved in signaling games, compositional programs, or VQA (Visual Question Answering), program generators and execution engines are updated jointly or sequentially with REINFORCE and cross-entropy losses under explicit bottleneck constraints (Vani et al., 2021, Carlsson et al., 2023).
• Transmission Bottleneck Implementation: Each generation samples a small "bottleneck" set $S$3 of training pairs; the student learns to generalize beyond $S$4 to reconstruct the full mapping. In semi-supervised variants, unsupervised batches of autoencoder reconstructions are interleaved with supervised training to enforce internal consistency and compositional bias.

A typical neural ILM training loop involves initial exposure to a bottleneck dataset, supervised updates to encoder and decoder networks, autoencoding updates on larger random samples, and thresholding or binarization before transmitting the language to the next generation (Bunyan et al., 2024, Lee et al., 6 Jan 2026).

3. Information Theory and the Bottleneck Principle

ILM dynamics are tightly coupled to ideas from rate-distortion theory and the Information Bottleneck (IB) framework. When languages are evaluated post-hoc on the IB plane—that is, by their mutual information $S$5 (complexity) and $S$6 (accuracy)—ILM-evolved systems lie close to the theoretical efficiency frontier (Carlsson et al., 2023). The transition bottleneck acts as a rate limiter: incremental increases in the bottleneck size (or communication channel rate $S$7) produce sharp transitions in system behavior, from degenerate performance to cumulative learning (Prystawski et al., 22 Nov 2025).

Information-theoretic analysis further implies:

• Cumulative culture ("ratcheting") only emerges when per-generation transmission exceeds a critical channel rate; below this, performance can plateau or even degrade ("misinformation" regime).
• Slight increases in per-generation communicative capacity can drive qualitative leaps in the outcomes of iterated learning, explaining the uniquely rapid, cumulative cultural evolution of human language (Prystawski et al., 22 Nov 2025).

4. Compositionality and the Emergence of Structure

Empirically, ILMs reliably induce compositional mapping structure even in neural agents lacking explicit architectural bias. Compositionality is typically quantified via:

• Topographic similarity $S$8, the correlation between distances in meaning space and signal space (Ren et al., 2020, Perkins, 2021);
• Information-theoretic metrics, such as conditional entropy between meaning and signal vector components;
• Holdout accuracy, measuring zero-shot generalization to unseen combinations (Perkins, 2021).

Iterated exposure to a bottlenecked sample leads agents to prefer languages where similar meanings are mapped to similar signals (high $S$9). This drift is robust—for sufficiently tight bottlenecks and repeated generations, neural ILMs produce nearly maximally compositional codes and outperform non-iterated baselines on systematic generalization tasks (Vani et al., 2021). Notably, empirical results confirm strong correlation between emergent compositionality and out-of-distribution performance (Vani et al., 2021, Ren et al., 2020).

In semi-supervised/neural models, compositional codes emerge spontaneously if internal reflection is incentivized through autoencoding losses; this effect scales to high-dimensional, noisily rendered inputs (e.g., seven-segment glyph images) (Lee et al., 6 Jan 2026).

5. Extensions: Populations, Language Contact, and LLM Chains

Population-level and contact dynamics extend ILM beyond the classic teacher-pupil chain:

• Population ILMs: Multiple agents interact on a communication network. Sufficient between-community interaction leads to global convergence; spatially embedded communities can sustain stable dialectal variation, impeding full language amalgamation unless cross-community interaction exceeds a critical fraction (Sains et al., 2023).
• Language Contact: In semi-supervised ILMs, the same transmission and bottleneck pressures that drive compositionality also protect grammatical core mappings during language mixing. With balanced input from two languages, novel mixed codes can emerge; with a small bias in exposure probability, one language typically re-emerges intact (Bullock et al., 2024).
• LLMs and Self-Improvement: The ILM framework conceptually unifies multi-round self-training (self-distill, self-instruct) and multi-agent self-play in LLMs. Bayesian ILM predicts bias amplification: any inductive bias present in the LLM prior will compound exponentially unless counteracted by a structured "interaction phase" (e.g., data augmentation, filtering, human feedback). Posterior entropy decreases over generations, and model outputs drift toward the prior's favored modes (Ren et al., 2024).

6. Empirical Phenomena, Scaling Laws, and Limitations

Quantitative studies substantiate key ILM predictions:

Property	Typical Observation	Reference
Expressivity plateau	$P(h \mid d) \propto P_0(h) \, P(d \mid h)$0, generations	(Lee et al., 6 Jan 2026, Bunyan et al., 2024)
Compositionality	$P(h \mid d) \propto P_0(h) \, P(d \mid h)$1	(Ren et al., 2020, Vani et al., 2021)
Stability	$P(h \mid d) \propto P_0(h) \, P(d \mid h)$2	(Bunyan et al., 2024, Lee et al., 6 Jan 2026)
Bottleneck law	$P(h \mid d) \propto P_0(h) \, P(d \mid h)$3	(Bunyan et al., 2024)
Nonlinear channel effect	Threshold at $P(h \mid d) \propto P_0(h) \, P(d \mid h)$4 bits	(Prystawski et al., 22 Nov 2025)

A critical result is the linear scaling law: the number of examples needed for iterated learning to induce stable, expressive, compositional languages grows only linearly in the semantic dimension, despite the mapping space growing exponentially. This yields computational tractability for large scale models (Bunyan et al., 2024).

Limitations of most current ILMs include absence of explicit population structure, lack of socio-cultural factors, and limited modeling of bilingualism and code-switching (Bullock et al., 2024). For neural ILMs, explicit architectural compositional biases (e.g., modularity) are often unnecessary, but strong autoencoding or internal reflection objectives become more critical with larger and noisier input spaces.

7. Broader Implications and Future Directions

ILM research provides a formal bridge between cultural evolution, cognitive science, information theory, and artificial intelligence:

• The bottleneck-induced drift toward compositionality, expressivity, and stability underlies the emergence of structure in natural and artificial languages.
• Information-theoretic thresholds explain regime transitions in cultural ratcheting and the unique evolutionary trajectory of human symbolic systems (Prystawski et al., 22 Nov 2025).
• In the domain of LLM evolution, ILM highlights the need for strong interaction phases or regulatory mechanisms to prevent excessive bias amplification and loss of diversity.

Future research directions include population-embedded ILM with overlapping generations, richer network topologies, complex meaning spaces (e.g., full images, multi-agent referential games), population dynamics in language contact, and explicit modeling of cognitive-level "language of thought" architectures in neural agents. Coupling ILM simulations with human and cross-species experiments, as well as scaling to realistic multimodal and multi-agent environments, are promising avenues for advancing the theory and applications of iterated learning.