[Paper Review] Symbolic Behaviour in Artificial Intelligence

Adam Santoro, Andrew K. Lampinen+3 more|arXiv (Cornell University)|Feb 5, 2021

Language and cultural evolution75 references4 citations

TL;DR

This paper proposes that symbolic fluency in AI should be evaluated not by symbolic architecture alone, but through behavioral criteria rooted in human socio-cultural interaction. By treating symbol meaning as conventional and interpreter-dependent, the authors introduce graded behavioral metrics—such as constructing new symbols, adapting conventions, and introspecting on meaning—and argue that learning-based agents can achieve human-like symbolic behavior when immersed in rich, interactive, human-interactive environments with feedback, optimizing behavior directly through scalable data collection.

ABSTRACT

The ability to use symbols is the pinnacle of human intelligence, but has yet to be fully replicated in machines. Here we argue that the path towards symbolically fluent artificial intelligence (AI) begins with a reinterpretation of what symbols are, how they come to exist, and how a system behaves when it uses them. We begin by offering an interpretation of symbols as entities whose meaning is established by convention. But crucially, something is a symbol only for those who demonstrably and actively participate in this convention. We then outline how this interpretation thematically unifies the behavioural traits humans exhibit when they use symbols. This motivates our proposal that the field place a greater emphasis on symbolic behaviour rather than particular computational mechanisms inspired by more restrictive interpretations of symbols. Finally, we suggest that AI research explore social and cultural engagement as a tool to develop the cognitive machinery necessary for symbolic behaviour to emerge. This approach will allow for AI to interpret something as symbolic on its own rather than simply manipulate things that are only symbols to human onlookers, and thus will ultimately lead to AI with more human-like symbolic fluency.

Motivation & Objective

• To reframe symbolic intelligence in AI not as a structural property but as a behavioral capability rooted in social convention.
• To address the limitations of binary symbol-use assessments by introducing graded, measurable criteria for symbolic behavior.
• To argue that symbolic fluency in machines emerges not from symbolic architectures per se, but from immersion in human socio-cultural interactions requiring coordination on meaning.
• To demonstrate that contemporary learning-based AI can be optimized for symbolic behavior through interactive, feedback-rich, large-scale data collection.

Proposed method

• Define symbols as triadic relations: a substrate, a denotation, and an interpreter, where meaning arises from agreed-upon convention.
• Frame symbolic behavior as a set of measurable, graded traits: receptiveness to new conventions, construction of new symbols, adaptation of existing symbols, and introspection on meaning and reasoning.
• Propose that symbolic fluency emerges not from symbolic machinery alone, but from agents learning in socio-cultural environments requiring coordination on arbitrary meanings.
• Advocate for collecting large-scale, interactive datasets of human behavior and incorporating human feedback to train agents to exhibit symbolic behaviors.
• Use modern learning algorithms to directly optimize agents for symbolic behaviors, treating symbolic fluency as a behavioral target rather than a structural one.
• Reject strict compositionality and discrete representations as necessary, emphasizing that continuous or distributed representations can still function symbolically if treated conventionally.

Experimental results

Research questions

• RQ1 How can symbolic behavior in AI be measured in a graded, non-binary way, rather than as a simple presence or absence?
• RQ2 What role do socio-cultural interactions and human feedback play in enabling symbolic fluency in learning-based agents?
• RQ3 Can symbolic behavior be optimized directly through behavioral objectives, even without explicit symbolic architectures?
• RQ4 To what extent is systematicity or compositional generalization necessary for symbolic behavior, and how does it emerge in practice?
• RQ5 How can continuous or distributed representations function as symbols if they are not discrete or localist?

Key findings

• Symbolic behavior is not a binary trait but a graded set of competencies, including constructing new symbols, adapting conventions, and introspecting on meaning.
• Human-like symbolic fluency in AI is not guaranteed by symbolic machinery alone but emerges from immersion in socio-cultural interactions that demand coordination on arbitrary meanings.
• Interactive, feedback-rich, and diverse experiences—especially those involving human collaboration—significantly enhance symbolic behavior beyond scale alone.
• Generalization and symbolic competence can emerge from rich, interactive experiences even without enforcing strict compositionality or discrete representations.
• The triadic nature of symbols—substrate, denotation, and interpreter—means that symbolic meaning is inherently subjective and convention-based, not intrinsic to the symbol itself.
• Optimizing learning-based agents directly for symbolic behaviors, using scalable data and human feedback, is a viable and effective path toward symbolic fluency.