Disentangling Linguistic Competence and Factual Knowledge in LLMs: A Survey

ACL ARR 2026 January Submission2883 Authors

Keywords: Large Language Models (LLMs) ; Linguistic–Knowledge Separation; Factuality and Hallucinations; Survey Paper

Abstract:

Maintaining factual accuracy is becoming increasingly expensive as Large Language Models (LLMs) scale. This has spurred a modular perspective that decouples linguistic competence from factual knowledge in LLMs, which enables targeted fact updates without full retraining. Yet a coherent synthesis to guide this emerging line of work is still lacking.
To fill this gap, we present a comprehensive survey through the lens of Linguistic–Knowledge Separation (LKS), consolidating methods and evaluations into a unified framework.
We make four contributions: (1) We clarify the conceptual distinction between linguistic and factual knowledge. (2) We summarize representative benchmarks and metrics for the linguistic and knowledge sides, enabling side-specific evaluation. (3) We provide a comprehensive summary of LKS methodologies and develop a systematic taxonomy that organizes them into coherent categories. (4) Finally, we outline future directions and open challenges toward robust, generalizable LKS.

Paper Type: Long

Research Area: Interpretability and Analysis of Models for NLP

Research Area Keywords: knowledge tracing/discovering/inducing; model editing; probing;

Contribution Types: Surveys

Languages Studied: English

Submission Number: 2883