Automated Semantic Rules Detection ( ASRD ) for Emergent Communication Interpretation

Bastien Vanderplaetse, Xavier Siebert, Stéphane Dupont · 2026

Read Automated Semantic Rules Detection ( ASRD ) for Emergent Communication Interpretation by Bastien Vanderplaetse, Xavier Siebert, Stéphane Dupont (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-03254.txt.

Introduction

Bastien Vanderplaetse, Xavier Siebert, and Stéphane Dupont help researchers interpret the private languages that AI agents invent. In emergent communication studies, agents develop their own signaling systems with no explicit programming. But few tools exist for working out what the messages mean. Their algorithm, ASRD, automatically detects semantic rules inside those message streams.

ASRD pulls out recurring patterns from messages exchanged by agents trained on two datasets for the Lewis signaling game, a standard testbed in the field. It then links the patterns to specific traits of the input data the agents were describing. By tying message fragments to concrete traits, the method makes later analysis much simpler. It gives researchers a systematic way to read an emergent language, instead of guessing at it.

Catalog record

CREATORBastien Vanderplaetse, Xavier Siebert, Stéphane Dupont
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-03254

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Automated Semantic Rules Detection ( ASRD ) for Emergent Communication Interpretation Bastien Vanderplaetse, Xavier Siebert, Stéphane Dupont Abstract The field of emergent communication within multi-agent systems examines how autonomous agents can independently develop communication strategies, without explicit programming, and adapt them to varied environments. However, few studies have focused on the interpretability of emergent languages. The research exposed in this paper proposes an Automated Semantic Rules Detection ( ASRD ) algorithm, which extracts relevant patterns in messages exchanged by agents trained with two different datasets on the Lewis Game, which is often studied in the context of emergent communication. ASRD helps at the interpretation of the emergent communication by relating the extracted patterns to specific attributes of the input data, thereby considerably simplifying subsequent analysis. 1 Introduction The study of emergent communication in multi-agent systems aims to uncover how autonomous agents can independently create and refine their own interaction protocols. This research area connects computational linguistics, artificial intelligence, and robotics, focusing on the development of complex communication strategies among autonomous agents. Depending on the task they are trained for, these agents may cooperate or be in competition with one another. To communicate, agents need a protocol communication, which refers to a set of rules, conventions, and structures that govern how information is exchanged between agents in a multi-agent system. These protocols define the format, timing, sequencing, and semantics of messages, ensuring that agents can effectively interpret and respond to the information they receive. More specifically, emergent communication protocols focus on the functional aspects of communication, such as how messages are encoded, transmitted, and decoded. These protocols govern the mechanics of interaction, ensuring that agents can exchange information efficiently and reliably. Communication protocols require a language, which refers to the symbolic system or vocabulary that agents develop to represent and convey meaning. It encompasses the creation of symbols, grammar, and shared understanding, allowing agents to express complex ideas and concepts. While emergent communication protocols deal with the "how" of communication, emergent language deals with the "what" and the shared understanding that emerges from interactions. In traditional systems, communication protocols and languages are often explicitly designed and predefined, tailored to specific tasks or environments. However, in the context of emergent communication, either the protocol, the language, or both are not specified, but instead evolve dynamically as agents interact and learn from their environment and each other. This emergent process allows for the development of flexible and adaptive communication strategies, which can be particularly advantageous in complex or novel scenarios where predefined protocols and languages may be insufficient or impractical. However, learned protocols and languages raise challenges in interpretability. This has driven interest in using discrete symbols and syntactic structures inspired by human language, valued for its adaptability and ability to convey intricate ideas across new contexts. Emergent communication thus proposes a solution to create scalable, adaptable discrete communication languages that could hopefully mirror human linguistic structures to improve interpretability. However, the interpretation of messages exchanged by the agents in the context of emergent communication has not yet been extensively studied. This paper proposes an interpretation method based on an algorithm, Automated Semantic Rules Detection ( ASRD ), that extracts patterns occuring in the emergent language between the agents. This algorithm identifies constant positions in messages by grouping them based on attribute or hyperattribute values representing the properties of the observed data. It then extracts representative patterns by removing global invariants, linking these patterns to the data properties. Our results show that ASRD is able to find relevant patterns in the messages used by agents and to determine their relation to attributes defining the input data observed by the agents. Section 2 reviews related works in the field of emergent communication, highlighting the gaps in interpretability and the need for automated analysis tools. Section 3 provides a detailed explanation of emergent communication, focusing on the fundamental concepts of vocabulary, messages, and language in multi-agent systems. Section 4 introduces the ASRD algorithm, describing its workflow and how it extracts semantic rules from emergent communication protocols. Section 5 outlines the experimental setup, including the agents’ architecture, image feature transformations, datasets, and evaluation metrics used in this paper. Section 6 presents the results of the experiments, analyzing the performance of the agents and the effectiveness of ASRD in extracting semantic rules from emergent languages. Section 7 discusses the implications of the findings, highlights the limitations of the current approach, and suggests directions for future research. 2 Related Works Emergent communication builds on the idea that, when agents share an environment and work towards common objectives, they can develop communication strategies without human intervention [ 5 ] . Recent advances have highlighted the flexibility of these emergent protocols, showcasing their ability to improve coordination and collaboration across diverse tasks and scenarios [ 14 , 18 ] . Current research primarily focuses on improving the overall performance of multi-agent systems, with additional emphasis on the resilience and adaptability of emergent protocols under various conditions [ 2 , 18 ] . While emergent protocols improve agent collaboration, a notable gap in the literature lies in the interpretability of emergent languages and their alignment with human understanding, which is a critical aspect for ensuring the usability and trust of such systems in real-world contexts [ 19 ] . Some research works propose to design adaptive languages that enhance the clarity and sparsity of messages to meet the needs of diverse human-agent teams [ 8 ] . Others works explore the development of communication strategies more aligned with human language [ 1 ] , specifically on its compositionality property [ 15 ] . However, the interpretation of language emerging from such techniques requires a lot of manual analysis. In this paper, we propose a method to help at this interpretation thereby minimizing manual analysis. To focus on the language emergence and its interpretation, we test our algorithm on a simple task involving two agents with a simple communication protocol: the Lewis Game, often used in studies of emergent communication [ 3 , 11 ] . 3 Emergent Communication Emergent communication in multi-agent systems ( MAS ) refers to the process by which autonomous agents develop their own communication protocol or language without explicit instructions. In this paper, we focus on emergent language. Therefore, this Section defines the fundamental concepts of an emergent language. 3.1 Vocabulary and Messages In MAS , communication between agents requires a vocabulary \mathcal{V} , which is a finite, non-empty set of symbols \left\{w_{1},\ldots,w_{n}\right\} , where n\in\mathbb{N}_{0} represents the number of symbols. Each element w_{i}\in\mathcal{V} is referred to as a word . The set of all possible finite sequences of words from \mathcal{V} is denoted as \mathcal{V}^{*} . Agents use the words from \mathcal{V} to construct messages. A message m=\left(w_{1},\ldots,w_{T}\right) is a finite sequence of words, where each w_{i}\in\mathcal{V} and T is the length of the message. The set of all possible messages is denoted as \mathcal{M}\subseteq\mathcal{V}^{*} . Within a message, each instance of a word from \mathcal{V} is called a token . 3.2 Language in Multi-Agent Systems A language \mathcal{L} over a vocabulary \mathcal{V} is a structured system of communication used by agents to convey information. Formally, a language is defined as a set of messages selected

Read the complete text