Algorithmic and Empirical contributions in Linguistic Architectures for Grounding Hallucinating Models

About this paper

This paper examines hallucination in large language models (LLMs) through the lens of linguistic grounding.Hallucinations-plausible yet inaccurate outputs-undermine reliability, interpretability, and trust in generative systems.Existing mitigation strategies, including retrieval-augmented generation, fact-checking, and reinforcement learning with human feedback, vary in effectiveness but share a reliance on post-hoc correction rather than representational grounding.By comparing algorithmic approaches that optimize model behavior with empirical methods that depend on observed or human-guided validation, this study reveals a structural gap: current systems lack a semantic foundation to constrain generative drift.To address this, the paper introduces linguistic frames-structured templates capturing meaning, roles, and contexts-as a pathway for embedding semantic constraints directly into model architectures.Framed grounding offers a route toward architectures that balance fluency with truthfulness, positioning semantic representation as central to sustainable hallucination mitigation.