LLM Self-Explanations as Design Material: Toward a Taxonomy

Paper i proceeding, 2026

Researchers and practitioners routinely manipulate properties of LLM self-explanations (their length, tone, confidence, structure) yet these design choices remain implicit and inconsistently named across studies. Without shared vocabulary, we cannot systematically compare findings, replicate studies, or accumulate knowledge about which explanation properties affect user outcomes. We present a preliminary taxonomy of six categories of self-explanation properties drawn from the XAI and HCI literature: surface, relational, structural, delivery, source, and semantic properties. For each category, we identify key properties with empirical grounding from studies that explicitly manipulated these properties. We offer this taxonomy as a starting point for more rigorous research on LLM self-explanation design and invite community feedback to refine it.