Risk Aware Benchmarking of Large Language Models

TL;DR

This paper introduces a statistical framework for benchmarking large language models by assessing socio-technical risks using stochastic dominance tests, enabling risk-aware model selection with quantifiable significance.

Contribution

It develops a novel distributional benchmarking method based on stochastic dominance, linking risk assessment to econometric and financial models for the first time in LLM evaluation.

Findings

Effective risk comparison of LLMs regarding instruction drift and toxicity.

Statistically significant differences identified among models.

Framework validated through theoretical analysis and empirical experiments.

Abstract

We propose a distributional framework for benchmarking socio-technical risks of foundation models with quantified statistical significance. Our approach hinges on a new statistical relative testing based on first and second order stochastic dominance of real random variables. We show that the second order statistics in this test are linked to mean-risk models commonly used in econometrics and mathematical finance to balance risk and utility when choosing between alternatives. Using this framework, we formally develop a risk-aware approach for foundation model selection given guardrails quantified by specified metrics. Inspired by portfolio optimization and selection theory in mathematical finance, we define a metrics portfolio for each model as a means to aggregate a collection of metrics, and perform model selection based on the stochastic dominance of these portfolios. The statistical significance of our tests is backed theoretically by an asymptotic analysis via central limit theorems instantiated in practice via a bootstrap variance estimate. We use our framework to compare various large language models regarding risks related to drifting from instructions and outputting toxic content.