Enhancing LLM Evaluations: The Garbling Trick

TL;DR

This paper introduces a method to transform existing LLM evaluation metrics into progressively more challenging tasks, revealing nuanced differences in reasoning abilities among models.

Contribution

The authors propose a general technique to enhance LLM evaluations by increasing task difficulty, uncovering performance distinctions not visible in standard assessments.

Findings

Enhanced evaluations highlight differences between base and reasoning LLMs.

New evaluation corpus demonstrates improved discrimination of model capabilities.

Method reveals reasoning skills beyond traditional metrics.

Abstract

As large language models (LLMs) become increasingly powerful, traditional evaluation metrics tend to saturate, making it challenging to distinguish between models. We propose a general method to transform existing LLM evaluations into a series of progressively more difficult tasks. These enhanced evaluations emphasize reasoning capabilities and can reveal relative performance differences that are not apparent in the original assessments. To demonstrate the effectiveness of our approach, we create a new multiple-choice test corpus, extend it into a family of evaluations, and assess a collection of LLMs. Our results offer insights into the comparative abilities of these models, particularly highlighting the differences between base LLMs and more recent "reasoning" models.