Abstract
Language relies on Theory of Mind (ToM), as language users must entertain and recognize communicative intentions. Conversely, ToM abilities profit from language, as language provides a means for expressing mental states explicitly. Given this interdependence, it has been hypothesized that language and ToM have co-evolved. This paper presents an agent-based model demonstrating that language and perspective-taking—a core component of ToM—likely co-evolved through cultural transmission. By simulating generations of agents, the research shows that a pressure to understand others' internal perspectives can drive the emergence of informative linguistic conventions, and vice versa.
Introduction: The "Chicken and Egg" of Cognition
In cognitive science, the relationship between language and Theory of Mind (ToM) presents a classic "chicken and egg" problem. To understand a speaker’s intent, a listener requires ToM to model the speaker's mental state. However, language is one of the primary tools humans use to learn about and express those very mental states. This suggests a deep interdependence where neither could have reached its modern complexity in isolation.
The authors address the hypothesis of cultural co-evolution: the idea that these two skills developed interdependently over generations. Specifically, the study investigates how agents can learn a lexicon (a set of word-to-object mappings) while simultaneously inferring a speaker's perspective (what the speaker finds salient in a given context) when neither the lexicon nor the perspective is directly observable.
Methods: Simulating Iterated Learning
The researchers utilized an agent-based model embedded within an iterated learning framework. In this paradigm, the linguistic output of one generation serves as the input for the next, simulating the cultural transmission of knowledge over time.
The Model Environment
Agents exist in a world where objects have varying levels of "salience" based on an agent's individual perspective. To communicate, a speaker selects a signal for a specific topic. A learner must then use Bayesian inference to solve a "double unknown" problem: they must simultaneously infer the speaker's lexicon and the speaker's internal perspective.
Agent Architectures
The simulation tested two distinct types of agents:
- Literal Agents: Speakers choose words based strictly on their internal lexicon. While listeners use perspective-taking to interpret signals, the speakers do not consider the listener's mental state when choosing a word.
- Pragmatic Agents: These agents utilize the Rational Speech Act (RSA) framework. Speakers actively optimize their utterances by predicting how a perspective-taking listener is likely to interpret them, creating a recursive loop of social reasoning.
Selection Pressures
The simulation evaluated two types of evolutionary pressure:
- Selection for Communicative Success: Rewards agents if the listener correctly identifies the intended object.
- Selection for Perspective-Inference: Rewards agents if the listener correctly guesses the speaker's internal perspective (salience levels).
Key Insights and Results
The Bootstrapping Miracle
The most striking finding is that learners can solve the "double unknown" problem. Despite not knowing the words or the speaker's perspective initially, agents can bootstrap one from the other using Bayesian inference, provided the language possesses a baseline level of informativeness. This suggests that the co-evolutionary path is computationally feasible even from a state of total ambiguity.
Mind-Reading as a Linguistic Driver
When the simulation selected specifically for the ability to infer perspectives, informative languages emerged as a byproduct. This indicates that the social necessity of "reading minds" may have been a primary driver for the cultural evolution of structured, informative language. Language, in this view, is a tool that evolved to make internal states more transparent.
Pragmatism as a Buffer
Pragmatic agents (RSA-based) proved significantly more robust than literal agents. They achieved high communicative success even when dealing with suboptimal or ambiguous languages. Because they reason about the listener’s mind during the act of speaking, they can navigate linguistic gaps that would otherwise lead to failure. Pragmatic agents reached high success under both communicative and perspective-inference selection pressures.
No Pressure, No Progress
Crucially, the model demonstrated that without explicit selection pressure, populations failed to establish linguistic conventions or successful perspective-inference. This highlights that the co-evolution of these traits is not an inevitable outcome of interaction, but a response to specific functional demands.
Discussion and Limitations
While the model provides a robust theoretical framework for co-evolution, it operates within a simplified environment. The "world" and "perspectives" are low-dimensional; scaling these findings to complex, multi-featured environments remains a challenge for future research. Additionally, the model assumes a clean transmission of language, whereas real-world cultural evolution involves noisy social structures and varying degrees of fidelity.
Ultimately, the results suggest that pragmatic reasoning is the "glue" that allows mind and language to evolve in tandem. This points toward a historical trajectory where early hominids may have developed basic RSA-style reasoning, which then facilitated the rapid expansion of both linguistic and social-cognitive capabilities.
Summary updated