Human-Aligned Faithfulness in Toxicity Explanations of LLMs
This work addresses the need for trustworthy LLMs in downstream tasks by focusing on toxicity explanation evaluation, though it is incremental as it builds on existing explainability research.
The paper tackles the problem of evaluating LLMs' reasoning about toxicity through their free-form explanations, proposing a novel multi-dimensional criterion called Human-Aligned Faithfulness (HAF) with six metrics, and finds that LLMs generate plausible explanations for simple prompts but break down in nuanced reasoning, leading to inconsistent and irrelevant responses.
The discourse around toxicity and LLMs in NLP largely revolves around detection tasks. This work shifts the focus to evaluating LLMs' reasoning about toxicity -- from their explanations that justify a stance -- to enhance their trustworthiness in downstream tasks. Despite extensive research on explainability, it is not straightforward to adopt existing methods to evaluate free-form toxicity explanation due to their over-reliance on input text perturbations, among other challenges. To account for these, we propose a novel, theoretically-grounded multi-dimensional criterion, Human-Aligned Faithfulness (HAF), that measures the extent to which LLMs' free-form toxicity explanations align with those of a rational human under ideal conditions. We develop six metrics, based on uncertainty quantification, to comprehensively evaluate HAF of LLMs' toxicity explanations with no human involvement, and highlight how "non-ideal" the explanations are. We conduct several experiments on three Llama models (of size up to 70B) and an 8B Ministral model on five diverse toxicity datasets. Our results show that while LLMs generate plausible explanations to simple prompts, their reasoning about toxicity breaks down when prompted about the nuanced relations between the complete set of reasons, the individual reasons, and their toxicity stances, resulting in inconsistent and irrelevant responses. We open-source our code at https://github.com/uofthcdslab/HAF and LLM-generated explanations at https://huggingface.co/collections/uofthcdslab/haf.