LLM の説明可能性に関するやさしい入門ガイド

**

# Introduction

AI Explainability (XAI) has dominated the real-world AI systems landscape over the past few years, with large language models (LLMs) being no exception. In these highly complex and powerful models, transitioning from static to dynamic evaluation becomes imperative to better understand how these black-box systems generate natural language outputs. In addition, synthesizing dynamic evaluation with robust statistical approaches and affordable, production-ready frameworks for observability are also pivotal trends under the radar in the industry.

This article discusses LLM explainability and outlines the advances, trends, and ongoing developments in this important field of study that attempts to measure, interpret, and better manage one of the most sophisticated forms of AI systems to date.

Even though LLMs have revolutionized the AI field as a whole, their inner workings remain largely opaque. High-stakes industries are increasingly turning to LLMs, deploying complex, specialized models where decisions made based upon their responses can have a significant impact. In this context, XAI, and more particularly LLM explainability, becomes more relevant than ever before.

The model's ability and "intelligence" to make decisions has been classically measured via public, static benchmarks. Yet recent studies suggest the traditional scorecard has broken down, with models' behavioral shift towards memorizing public tests instead of proving true reasoning. The need for dynamic, multidimensional evaluation frameworks has significantly arisen: these frameworks evaluate systems against novel scenarios grounded by experts.

But what does XAI really seek beyond merely evaluating whether an LLM is correct or incorrect in its responses? It primarily seeks to understand *why*. In this sense, model-agnostic local explanations constitute an effective approach, with state-of-the-art frameworks like SMILE**-based ones — SMILE being an acronym for Statistical Model-Agnostic Interpretability with Local Explanations — that analyze the impact of slight alterations in user prompts (model inputs) on the resulting generated text. These frameworks do not limit themselves to using basic proximity measurements. Instead, they apply advanced, rigorous statistical distance measures. As a result, they can build robust artifacts like visual heatmaps that pinpoint which parts of the input (e.g. words) were most influential in the model's decision to generate a certain output.

The following diagram shows how to address the issue of little or no model transparency. gSMILE, a framework based on SMILE, can be used to explain how LLMs respond to different parts of a prompt.

**

gSMILE explains how LLMs provide responses to distinct parts of a prompt | Image by LLM-SMILE

Having these cutting-edge frameworks for evaluating LLMs' internal reasoning may sound fantastic at first glance. However, building local, prompt-wise explanations can easily become prohibitive when it comes to massive, closed-source LLMs, as these models manage a huge volume of API calls. This motivated the need for solutions that are accessible and budget-friendly, as pointed out in recent studies. In this direction, researchers have built a proxy solution that employs smaller, open-source models as a means to approximate and simplify the otherwise complex decision boundaries of proprietary LLMs. Their mechanism ensures high-fidelity explanations as costs are significantly reduced, which makes model interpretability accessible even for everyday developers.

Beyond theoretical and scientific progress, there are increasing shifts towards practical observability, with engineering relying on tracking platforms such as CometLLM**. These frameworks, envisioned to democratize explainability, can capture prompt iterations, granular metadata, and traces of previous executions. Consequently, developers gain the ability to debug pipelines and make workflows reproducible, all without the need for a deep mathematical understanding.

# Summing Up

**

The progress and prospects analyzed lead us to conclude that the vast ecosystem of LLM XAI is rapidly accelerating. Amid this explosion of research and the appearance of free-friendly solutions, community-driven hubs for LLM XAI are becoming essential. A combination of robust statistical evaluation with engineering approaches positioned on the budget-friendly side of the spectrum is key to gradually opening the black box and promoting models that are not only powerful, but also trustworthy and transparent.

Key references, for further reading:

• Awesome-LLM-Explainability (GitHub Repository)
• R. Olson. 2025 Year in Review for LLM Evaluation: When the Scorecard Broke, Goodeye Labs, 2025.
• J. Liu, et al. Revitalizing Black-Box Interpretability: Actionable Interpretability for LLMs via Proxy Models (arXiv).
• LLM-SMILE (GitHub Repository)
• S. Tripathi. A Hands-on Guide on CometLLM for LLM Explainability. ADaSci, 2024.

Iván Palomares Carrascosa** is a leader, writer, speaker, and adviser in AI, machine learning, deep learning & LLMs. He trains and guides others in harnessing AI in the real world.