How to systematically and quantifiably remove meaning?

Area of Science:

• Artificial Intelligence
• Computational Linguistics
• Cognitive Science

Background:

• Large language models (LLMs) are increasingly used in real-world applications.
• Current methods for evaluating LLM robustness against input meaning degradation are insufficient.
• A systematic approach is needed to quantify performance decline due to semantic erosion.

Purpose of the Study:

• To develop and validate a framework for semantically eroding input meaning.
• To quantify the intensity of semantic erosion and its impact on LLM performance.
• To identify domain-specific vulnerabilities in LLM processing.

Main Methods:

• Developed five theoretically motivated semantic erosion methods: omission, lexical substitution, abstraction, structural obfuscation, and logical error injection.
• Applied erosion operators across five distinct domains (e.g., code generation, news, instructions).
• Quantified LLM performance degradation using a publicly available model and two-way Analysis of Variance (ANOVA).

Main Results:

• Significant main effects of both domain and erosion method on LLM performance were observed.
• A significant interaction effect indicated that semantic degradation impact is dependent on both erosion type and domain-specific information.
• Logical errors severely impacted code generation, while structural obfuscation most affected news and instruction tasks. Pairwise erosion combinations showed both synergistic and compensatory effects.

Conclusions:

• LLM performance degradation is highly sensitive to the type of semantic erosion and the specific domain.
• Domain-specific vulnerability profiles are crucial for robust LLM evaluation, moving beyond generic perturbations.
• Semantic erosion provides a principled method for analyzing how LLMs process and degrade meaning, aiding in understanding model failure modes.