The XAI-Seeking Principle: Structuring Explainability for LLM-Mediated Decision Support
TL;DR
The XAI-Seeking Principle (XSP) is introduced, a hypertextual design principle that structures explainability as linked abstraction layers: overview-links connect low-level evidence to higher-level summaries, while detail-links preserve inspectable evidence.
Abstract
We investigate how LLM-mediated explanations can preserve evidence trails, support orientation, and reduce interpretation effort under cognitive and temporal constraints. While LLMs can make XAI artifacts accessible, fluent summaries may obscure provenance and hinder verification. We introduce the XAI-Seeking Principle (XSP), a hypertextual design principle that structures explainability as linked abstraction layers: overview-links connect low-level evidence to higher-level summaries, while detail-links preserve inspectable evidence. Higher levels support rapid action; lower levels enable justification and verification. XSP guides transformation monitoring, feature alignment, performance–latency trade-offs, and validation-driven refinement. We examine it in a misinformation-support prototype for social media posts. Results reveal layer-specific effects of model scale and reasoning, trade-offs between semantic grounding and assessment stability, and the value of validation signals for adapting transformations. XSP thus provides a design principle for constructing explainability as a navigable representation space, with technical evidence on how abstraction layers, feature grounding, latency, and validation interact in LLM-mediated XAI.