Vintage Code, Modern Judges: Meta-Validation in Low Data Regimes

TL;DR

This work tackles the problem of validating LLM-based evaluators (LaaJs) in low-data regimes for legacy-code explanations, where labeled data are scarce and traditional evaluation is unreliable. It introduces SparseAlign, consisting of two components: Human-Rank to derive a consensus model ranking under sparse, disjoint human judgments using a delta-threshold tie-clustering and weighted pairwise voting with confidence, and Align-Score to quantify agreement via a confidence-weighted rank discrepancy and score proximity term, with align-score defined as $\mathit{align\text{-}score} = 1 - \varepsilon$ and $\varepsilon = \alpha \epsilon_{\text{rank}} + (1 - \alpha) \epsilon_{\text{score}}$. The method is validated on a COBOL code-explanation task with 7 human evaluators and 6 LaaJ candidates, showing SparseAlign can distinguish high-quality evaluators from weak ones and identify top-performing LaaJs, thereby informing safe production deployment. Overall, SparseAlign provides a principled, replicable framework for validating LaaJ evaluators in settings with scarce annotated data, enabling more reliable, high-stakes decisions in legacy-system modernization.

Abstract

Application modernization in legacy languages such as COBOL, PL/I, and REXX faces an acute shortage of resources, both in expert availability and in high-quality human evaluation data. While Large Language Models as a Judge (LaaJ) offer a scalable alternative to expert review, their reliability must be validated before being trusted in high-stakes workflows. Without principled validation, organizations risk a circular evaluation loop, where unverified LaaJs are used to assess model outputs, potentially reinforcing unreliable judgments and compromising downstream deployment decisions. Although various automated approaches to validating LaaJs have been proposed, alignment with human judgment remains a widely used and conceptually grounded validation strategy. In many real-world domains, the availability of human-labeled evaluation data is severely limited, making it difficult to assess how well a LaaJ aligns with human judgment. We introduce SparseAlign, a formal framework for assessing LaaJ alignment with sparse human-labeled data. SparseAlign combines a novel pairwise-confidence concept with a score-sensitive alignment metric that jointly capture ranking consistency and score proximity, enabling reliable evaluator selection even when traditional statistical methods are ineffective due to limited annotated examples. SparseAlign was applied internally to select LaaJs for COBOL code explanation. The top-aligned evaluators were integrated into assessment workflows, guiding model release decisions. We present a case study of four LaaJs to demonstrate SparseAlign's utility in real-world evaluation scenarios.

Figures (2)

• Figure 1: Human evaluation on COBOL: Six models explained the same 25 samples; seven annotators received sparse, disjoint subsets.
• Figure 2: SparseAlign correctly assigns low alignment to the weak evaluator $L_{\mathrm{random}}$ and high alignment to the strong evaluator $L_{\mathrm{human\_close}}$, while quantifying the performance of the two real LaaJs.