Evaluating LLMs' Reasoning Over Ordered Procedural Steps

TL;DR

This paper assesses the ability of large language models to reconstruct ordered procedural sequences from shuffled steps, revealing their limitations especially with longer and more disordered inputs, using a new evaluation framework and dataset.

Contribution

The study introduces a comprehensive evaluation framework with adapted metrics for assessing LLMs' procedural reasoning over ordered sequences, and provides empirical insights into their performance limitations.

Findings

Model performance decreases with longer sequences.

Greater step shuffling worsens model accuracy.

Current LLMs struggle with complex procedural ordering.

Abstract

Reasoning over procedural sequences, where the order of steps directly impacts outcomes, is a critical capability for large language models (LLMs). In this work, we study the task of reconstructing globally ordered sequences from shuffled procedural steps, using a curated dataset of food recipes, a domain where correct sequencing is essential for task success. We evaluate several LLMs under zero-shot and few-shot settings and present a comprehensive evaluation framework that adapts established metrics from ranking and sequence alignment. These include Kendall's Tau, Normalized Longest Common Subsequence (NLCS), and Normalized Edit Distance (NED), which capture complementary aspects of ordering quality. Our analysis shows that model performance declines with increasing sequence length, reflecting the added complexity of longer procedures. We also find that greater step displacement in the input, corresponding to more severe shuffling, leads to further degradation. These findings highlight the limitations of current LLMs in procedural reasoning, especially with longer and more disordered inputs.