Explainability and Interpretability Methods for Large Language Models: An Exploratory Systematic Mapping Study
Year: 2026
Type: preprint
Abstract: Explainability for large language models (LLMs) spans circuit analysis, feature discovery, attribution, probing and editing, rationale faithfulness, source attribution, benchmarks and human-facing explanation. This systematic mapping study charts studies that explain an LLM or an LLM-centred pipeline, excluding work that utilises an LLM to explain another system. Seven sources were searched for reports from 2019 to August 2026 that explain at least one decoder-only or encoder–decoder model of at least 1B parameters. Of 7,825 records (5,730 after deduplication), 1,252 studies were included and coded by method family, explanandum, required access, evaluation construct and a four-field evidence profile. Screening, eligibility and coding were performed by a disclosed AI reviewer whose every decision carries a verbatim quote, and were validated against a human reviewer: title/abstract agreement reached κ = 0.84 after two failed calibration rounds, full-text agreement κ = 0.72, and conservative screening recall was 0.78. Methods requiring weights or activations account for 69.8% of the studies and methods compatible with hosted text or logit access for 23.3%; faithfulness or causal-validity evidence is coded in 70.4% (an approximate AI-coded estimate), human utility in 3.4%, deployed settings in 7 studies and specific regulatory obligations in 10. The map indicates that most explanation methods require access that hosted LLMs do not expose, and that human-centred and deployment evidence remains scarce. The protocol, its amendment and every decision are released.
Source: Research Square
Institution: University of Piraeus
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225664430
Cites:
Cited by:
Related to: