Examining the robustness of LLM evaluation to the distributional assumptions of benchmarks

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Ailem, Melissa, Marazopoulou, Katerina, Siska, Charlotte, Bono, James
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909218044379136
author Ailem, Melissa
Marazopoulou, Katerina
Siska, Charlotte
Bono, James
author_facet Ailem, Melissa
Marazopoulou, Katerina
Siska, Charlotte
Bono, James
contents Benchmarks have emerged as the central approach for evaluating Large Language Models (LLMs). The research community often relies on a model's average performance across the test prompts of a benchmark to evaluate the model's performance. This is consistent with the assumption that the test prompts within a benchmark represent a random sample from a real-world distribution of interest. We note that this is generally not the case; instead, we hold that the distribution of interest varies according to the specific use case. We find that (1) the correlation in model performance across test prompts is non-random, (2) accounting for correlations across test prompts can change model rankings on major benchmarks, (3) explanatory factors for these correlations include semantic similarity and common LLM failure points.
format Preprint
id arxiv_https___arxiv_org_abs_2404_16966
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Examining the robustness of LLM evaluation to the distributional assumptions of benchmarks
Ailem, Melissa
Marazopoulou, Katerina
Siska, Charlotte
Bono, James
Computation and Language
Benchmarks have emerged as the central approach for evaluating Large Language Models (LLMs). The research community often relies on a model's average performance across the test prompts of a benchmark to evaluate the model's performance. This is consistent with the assumption that the test prompts within a benchmark represent a random sample from a real-world distribution of interest. We note that this is generally not the case; instead, we hold that the distribution of interest varies according to the specific use case. We find that (1) the correlation in model performance across test prompts is non-random, (2) accounting for correlations across test prompts can change model rankings on major benchmarks, (3) explanatory factors for these correlations include semantic similarity and common LLM failure points.
title Examining the robustness of LLM evaluation to the distributional assumptions of benchmarks
topic Computation and Language
url https://arxiv.org/abs/2404.16966