Beyond Tokens: Introducing Large Semiosis Models (LSMs) forGrounded Meaning in Artificial Intelligence

Source abstract

Large Language Models (LLMs) represent a significant leap in artificial intelligence, demon-strating remarkable capabilities in processing and generating human-like text. However, their op-erational paradigm, primarily based on statistical correlations between symbolic tokens (signifier-s/representamens), revealsfundamentallimitationsconcerninggenuineunderstandingandsemanticgrounding. This paper posits that semiotic theory, drawing upon the foundational frameworks ofFerdinand de Saussure and Charles Sanders Peirce, offers essential analytical tools for diagnos-ing these deficiencies and proposing advancements. We argue that LLMs predominantly modelthe Saussurean signifier or the Peircean representamen, remaining largely disconnected from theconceptual signified or the referential object and meaning-effect interpretant. To address this crit-ical semantic gap, we introduce the concept of Large Semiosis Models (LSMs). LSMs areconceived as next-generation AI systems architected to explicitly model the triadic or dyadic rela-tionships inherent in sign processes, thereby integrating representations of meaning and referencewith symbolic manipulation. This paper outlines the theoretical rationale for LSMs, delineatestheir potential capabilities—including enhanced reasoning, robust grounding, and meaningful in-teraction—and proposes distinct implementation strategies inspired by Saussurean and Peirceansemiotics. Conceptual Python implementations using the LangChain framework are sketched toillustrate pathways for adapting current technologies towards LSM development. We concludethat the pursuit of LSMs constitutes a vital research trajectory for fostering AI systems exhibitinggreater robustness, reliability, and semantic intelligence.