Forgetting Our Way to Shared Meaning Effects of Forgetting on Conceptual Alignment in a Non-Partnership Coordination Game

Jul 16, 2026

Upload by Ophelia

Shared meaning in language depends on agents learning and agreeing on conceptual categories. This work studies how variations in agents’ memory affect the emergence and evolution of shared meaning. It models conceptual alignment as a non-partnership coordination game and compares actual versus perceived conceptual convergence using counterfactual simulations with different adaptiveness and memory degradation. Adaptive agents converge faster with closer final conceptual regions, whereas non-adaptive agents show earlier perceived convergence. Reducing the influence of novel information over time yields more stable agreements; memory properties are decisive for both actual and perceived alignment.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages7

Reading Time18 min

Abstract

Introduction

Semantic Memory and Large Language Models

Conceptual Semantics and Conceptual Regions

Forgetting Our Way to Shared Meaning: Effects of Forgetting on Conceptual Alignment in a Non-Partnership Coordination Game Landon C.L. Liu ([lcliu@andrew.cmu.edu](lcliu@andrew.cmu.edu)) 1 , Alan Tsang ( AlanTsang@cunet.carleton.ca) 3 & Mary Kelly (mary.kelly4@carleton.ca)4 1Department of Philosophy, Carnegie Mellon University 1Institute for Complex Social Dynamics, Carnegie Mellon University 3Department of Computer Science, Carleton University 4Department of Cognitive Science, Carleton University arXiv :2607 . 11787v1 [ cs .MA] 13 Jul 2026 Abstract Shared meaning in language requires people to learn and agree on categories. We ask how characteristics of agents’ memories change the emergence and evolution of shared meaning. Without a coordination game, models of conceptual semantics cannot explain how shared meaning emerges and changes in groups of people; however, existing games assume that players share payoffs in a partnership setting. We model conceptual alignment as a non-partnership game and illustrate differences in actual and perceived conceptual convergence from counterfactual simulations using agents with varying levels of adaptiveness and memory degradation. We found that adaptive players achieved actual convergence faster and had closer final conceptual regions than non-adaptive players, while non-adaptive players perceived convergence earlier. Weighing novel information less over time resulted in more stable agreements than fixing the weight of novel information. Memory features are critical to the emergence and evolution of actual and perceived convergence. Keywords: Concepts; Semantic Memory; Agent-based Modeling; Evolution of Language; Conceptual Semantics Introduction “How has your week been?” is often followed by, “It’s been good.” Politeness aside, even with an earnest and forthcoming friend, what constitutes “a good week” can vary remarkably depending on the friend. The concept of “goodness” here, localized to just events that can possibly happen to a person in a given week, is not a concept that suffers from a lack of data. Hopefully, most of us are being asked this question on a regular basis. We have no shortage of relevant data from regular use. Yet, it can be difficult to confidently feel that we always understand what a good week means–for some, that means successfully evading devastating calamity. For others, it may require achieving brilliance. The concept of “a good week,” as opposed to, say,“a bad week,” can be thought of as a partition of the concept of a week, with some decision boundary between good and bad. What we are interested in most broadly is how partitions converge in similarity as shared meanings develop, and how shared meanings develop based cognitive features of our language learning capacities. We are interested in how features of semantic memory (i.e., interpretable as rates of forgetting, levels of entrenchment, or reduced/a continued reduction in plasticity) affect the quality, the stability, and the speed of our shared agreements on these partitions. What can an architecture that allows for semantic memory tell us about misalignment, premature agreement and fragmentation in real linguistic communities? Semantic memory is a heavily studied topic, especially with the rise in the use of and the research into Large Language Models (LLMs; Bubeck et al., 2023; Vaswani et al., 2017); however, LLMs face critical theoretical and practical problems. First, LLMs face what has been called the symbol grounding problem (Harnad, 1990); it is not clear how meaning is represented or even extracted from the training datasets of LLMs, or that it is indeed “grounded”. Second, training LLMs is informationally expensive when compared to empirical observations of human learning. Humans, particularly young children, can learn new words and their meanings remarkably quickly; for example, children only need a few instances of a word to learn its meaning; far fewer than the larg

FAQs

What problem does the study address about shared meaning in language?

The study examines how shared meaning emerges and evolves when agents must learn and agree on categories, and how memory characteristics influence that process.

How is conceptual alignment modeled in this work?

Conceptual alignment is modeled as a non-partnership coordination game, rather than assuming players share payoffs in a partnership setting.

What effects of memory and adaptiveness does the study find?

Adaptive agents reach actual convergence faster and end with closer conceptual regions, while non-adaptive agents perceive convergence earlier. Memory features also strongly shape both actual and perceived convergence.

Jul 16, 2026

Upload by Ophelia