Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Metamorphic Testing of Large Language Models for Natural Language Processing

Steven Cho, Stefano Ruberto, Valerio Terragni

TL;DR

Metamorphic Testing (MT) provides an oracle-free mechanism to detect faulty NLP outputs from Large Language Models (LLMs). The authors compile the largest metamorphic-relations (MR) catalog for NLP (191 MRs) and introduce LLMorph, an automated MT framework implementing 36 MRs to generate and evaluate 561k metamorphic test groups across four NLP tasks and three LLMs. Findings show MT detects faults in about 18% of test groups and complements traditional labeled-data testing by catching faults not present in ground-truth labels, though false positives from input transformations and semantic comparisons remain a key challenge. The work delivers a valuable MR resource and an extensible testing framework that can be integrated into deployment pipelines, while outlining concrete directions to improve transformation quality and fault discrimination for practical AI reliability improvements.

Abstract

Using large language models (LLMs) to perform natural language processing (NLP) tasks has become increasingly pervasive in recent times. The versatile nature of LLMs makes them applicable to a wide range of such tasks. While the performance of recent LLMs is generally outstanding, several studies have shown that they can often produce incorrect results. Automatically identifying these faulty behaviors is extremely useful for improving the effectiveness of LLMs. One obstacle to this is the limited availability of labeled datasets, which necessitates an oracle to determine the correctness of LLM behaviors. Metamorphic testing (MT) is a popular testing approach that alleviates this oracle problem. At the core of MT are metamorphic relations (MRs), which define relationships between the outputs of related inputs. MT can expose faulty behaviors without the need for explicit oracles (e.g., labeled datasets). This paper presents the most comprehensive study of MT for LLMs to date. We conducted a literature review and collected 191 MRs for NLP tasks. We implemented a representative subset (36 MRs) to conduct a series of experiments with three popular LLMs, running approximately 560,000 metamorphic tests. The results shed light on the capabilities and opportunities of MT for LLMs, as well as its limitations.

Metamorphic Testing of Large Language Models for Natural Language Processing

TL;DR

Metamorphic Testing (MT) provides an oracle-free mechanism to detect faulty NLP outputs from Large Language Models (LLMs). The authors compile the largest metamorphic-relations (MR) catalog for NLP (191 MRs) and introduce LLMorph, an automated MT framework implementing 36 MRs to generate and evaluate 561k metamorphic test groups across four NLP tasks and three LLMs. Findings show MT detects faults in about 18% of test groups and complements traditional labeled-data testing by catching faults not present in ground-truth labels, though false positives from input transformations and semantic comparisons remain a key challenge. The work delivers a valuable MR resource and an extensible testing framework that can be integrated into deployment pipelines, while outlining concrete directions to improve transformation quality and fault discrimination for practical AI reliability improvements.

Abstract

Using large language models (LLMs) to perform natural language processing (NLP) tasks has become increasingly pervasive in recent times. The versatile nature of LLMs makes them applicable to a wide range of such tasks. While the performance of recent LLMs is generally outstanding, several studies have shown that they can often produce incorrect results. Automatically identifying these faulty behaviors is extremely useful for improving the effectiveness of LLMs. One obstacle to this is the limited availability of labeled datasets, which necessitates an oracle to determine the correctness of LLM behaviors. Metamorphic testing (MT) is a popular testing approach that alleviates this oracle problem. At the core of MT are metamorphic relations (MRs), which define relationships between the outputs of related inputs. MT can expose faulty behaviors without the need for explicit oracles (e.g., labeled datasets). This paper presents the most comprehensive study of MT for LLMs to date. We conducted a literature review and collected 191 MRs for NLP tasks. We implemented a representative subset (36 MRs) to conduct a series of experiments with three popular LLMs, running approximately 560,000 metamorphic tests. The results shed light on the capabilities and opportunities of MT for LLMs, as well as its limitations.

Paper Structure

This paper contains 11 sections, 1 equation, 3 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Metamorphic Testing for LLMs
  3. Literature Search of MRs
  4. Experiments
  5. RQ1: Failure Rate
  6. RQ2: Comparison with Traditional Testing
  7. RQ3 - Manual Validation
  8. Discussion and Analysis
  9. Threats to validity
  10. Related Work
  11. Conclusion and Future Work

Figures (3)

  • Figure 1: Real faulty behaviour detected by LLMorph in the OpenAI gpt-4-1106 model.
  • Figure 2: Number of the 44 papers published each year from 2018 to 2024
  • Figure 3: Stacked bar chart of the manual labels by LLM, Task, and MR (RQ3)

Theorems & Definitions (3)

  • Definition 1
  • Definition 2
  • Definition 3