Evaluating large language model (LLM) outputs in the legal domain presents unique challenges due to the complex and nuanced nature of legal analysis. Current evaluation approaches either depend on reference data, which is costly to produce, or use standardized assessment methods, both of which have significant limitations for legal applications.Although LLM-as-a-Judge has emerged as a promising evaluation technique, its reliability and effectiveness in legal contexts depend heavily on evaluation processes unique to the legal industry and how trustworthy the evaluation appears to the human legal expert. This is where existing evaluation methods currently fail and exhibit considerable variability.This paper aims to close the gap: a) we break down lengthy responses into “Legal Data Points” (LDPs) — self-contained units of information — and introduce a novel, reference-free evaluation methodology that reflects how lawyers evaluate legal answers; b) we demonstrate that our method outperforms a variety of baselines on both our proprietary dataset and an open-source dataset (LegalBench); c) we show how our method correlates more closely with human expert evaluations and helps improve inter-annotator agreement; and finally d) we open source our Legal Data Points for a subset of LegalBench used in our experiments, allowing the research community to replicate our results and advance research in this vital area of LLM evaluation on legal question-answering.

Several explanation methods such as Integrated Gradients (IG) can be characterised as path-based methods, as they rely on a straight line between the data and an uninformative baseline. However, when applied to language models, these methods produce a path for each word of a sentence simultaneously, which could lead to creating sentences from interpolated words either having no clear meaning, or having a significantly different meaning compared to the original sentence. In order to keep the meaning of these sentences as close as possible to the original one, we propose Sequential Integrated Gradients (SIG), which computes the importance of each word in a sentence by keeping fixed every other words, only creating interpolations between the baseline and the word of interest. Moreover, inspired by the training procedure of language models, we also propose to replace the baseline token “pad” with the trained token “mask”. While being a simple improvement over the original IG method, we show on various models and datasets that SIG proves to be a very effective method for explaining language models.