Unearthing Skill-Level Insights for Understanding Trade-Offs of Foundation Models

TL;DR

This paper introduces an automatic method to analyze model performance on specific skills within benchmarks, revealing nuanced trade-offs and enabling targeted improvements by inspecting model-generated rationales.

Contribution

The authors propose a novel skill-recovery approach using rationales, enabling detailed skill-wise evaluation across multiple benchmarks and models, which was previously obscured by aggregate metrics.

Findings

Identified common skills across benchmarks, creating skill-slices for detailed analysis.

Revealed significant performance trade-offs between models on specific skills.

Demonstrated improved accuracy by routing instances based on skill strengths.

Abstract

With models getting stronger, evaluations have grown more complex, testing multiple skills in one benchmark and even in the same instance at once. However, skill-wise performance is obscured when inspecting aggregate accuracy, under-utilizing the rich signal modern benchmarks contain. We propose an automatic approach to recover the underlying skills relevant for any evaluation instance, by way of inspecting model-generated rationales. After validating the relevance of rationale-parsed skills and inferring skills for $46$k instances over $12$ benchmarks, we observe many skills to be common across benchmarks, resulting in the curation of hundreds of skill-slices (i.e. sets of instances testing a common skill). Inspecting accuracy over these slices yields novel insights on model trade-offs: e.g., compared to GPT-4o and Claude 3.5 Sonnet, on average, Gemini 1.5 Pro is $18\%$ more accurate in "computing molar mass", but $19\%$ less accurate in "applying constitutional law", despite the overall accuracies of the three models differing by a mere $0.4\%$. Furthermore, we demonstrate the practical utility of our approach by showing that insights derived from skill slice analysis can generalize to held-out instances: when routing each instance to the model strongest on the relevant skills, we see a $3\%$ accuracy improvement over our $12$ dataset corpus. Our skill-slices and framework open a new avenue in model evaluation, leveraging skill-specific analyses to unlock a more granular and actionable understanding of model capabilities.