PaperlayerView exact revision on arXiv

Source-linked AI summary

How Robust is GPT-3.5 to Predecessors? A Comprehensive Study on Language Understanding Tasks

Xuanting Chen, Junjie Ye, Can Zu, Nuo Xu, Rui Zheng, Minlong Peng, Jie Zhou, Tao Gui, Qi Zhang, Xuanjing Huang

arXiv:2303.00293v1cs.CL

TL;DR

Robustness of large language models remains under-studied despite its relevance to open-world complexity and trustworthy AI. This study evaluates GPT-3.5 across NLU datasets and transformations, finding task-specific strengths alongside substantial robustness degradation and other challenges.

  • Problem

    Despite studies of LLM performance across tasks, their robustness to open-world complexities remains less examined, although it is relevant to trustworthy AI.

    Source: §1 · p. 1, §2.1 · p. 4

  • Method

    The study evaluates GPT-3.5 using 21 datasets across nine NLU tasks, 66 TextFlint transformations, and zero- and few-shot settings.

    Source: §1 · p. 2, §3.4 · p. 5

  • Results

    Across 21 datasets and nine NLU tasks, GPT-3.5 outperforms fine-tuned models on some tasks but shows robustness degradation, including drops of up to 35.74% in natural language inference and 43.59% in sentiment analysis.

    Source: §1 · p. 2, §5 · p. 19

  • Takeaways & Limitations

    The authors identify robustness instability, prompt sensitivity, and number sensitivity as challenges for GPT-3.5 and call for further research into robustness and prompting.

    Source: §1 · p. 2, §5 · p. 19

Abstract

from arXiv · show

The GPT-3.5 models have demonstrated impressive performance in various Natural Language Processing (NLP) tasks, showcasing their strong understanding and reasoning capabilities. However, their robustness and abilities to handle various complexities of the open world have yet to be explored, which is especially crucial in assessing the stability of models and is a key aspect of trustworthy AI. In this study, we perform a comprehensive experimental analysis of GPT-3.5, exploring its robustness using 21 datasets (about 116K test samples) with 66 text transformations from TextFlint that cover 9 popular Natural Language Understanding (NLU) tasks. Our findings indicate that while GPT-3.5 outperforms existing fine-tuned models on some tasks, it still encounters significant robustness degradation, such as its average performance dropping by up to 35.74\% and 43.59\% in natural language inference and sentiment analysis tasks, respectively. We also show that GPT-3.5 faces some specific robustness challenges, including robustness instability, prompt sensitivity, and number sensitivity. These insights are valuable for understanding its limitations and guiding future research in addressing these challenges to enhance GPT-3.5's overall performance and generalization abilities.

1 Introduction

The introduction frames GPT-3.5 robustness as an under-studied dimension of real-world reliability and evaluates it across NLU tasks. Results show both competitive task performance and substantial, task-dependent robustness challenges.

1 Introduction

  • The study evaluates GPT-3.5 on 21 datasets and 115,715 test samples across nine NLU tasks, using 66 TextFlint transformations and comparing models and prompting settings.The evaluation also examines specific robustness challenges and differences between GPT-3 and GPT-3.5 series models.Source: §1 · p. 2
  • GPT-3.5 achieves state-of-the-art results against task-specific fine-tuned models on some NLU tasks, with strengths in reading comprehension and sentiment analysis and challenges in sequence tagging and relation extraction.Figure 1 compares zero- and few-shot GPT-3.5 results with BERT-base and task-fine-tuned models across 21 datasets, showing the better result among prompts for each dataset.Source: §1 · p. 2, Figure 1 · p. 3
  • Robustness drops by up to 35.74% in natural language inference and 43.59% in sentiment analysis, while reading comprehension and WSC show notable robustness.GPT-3.5 also performs well on reading comprehension and sentiment analysis, but faces challenges in sequence tagging and relation extraction.Source: §1 · p. 2, Figure 1 · p. 3
  • In few-shot settings, robustness varies by task: it improves in aspect-based sentiment analysis but decreases in natural language inference and semantic matching.The introduction calls this cross-task variation robustness instability.Source: §1 · p. 2
  • Prompt changes substantially affect results, with high variance across sentiment analysis, natural language inference, semantic matching, and reading comprehension.The introduction says robustness to prompt variations still requires improvement.Source: §1 · p. 2
  • GPT-3.5 is more sensitive to numerical inputs than pre-training fine-tuning models, particularly when numerical words are replaced with different values.This sensitivity is illustrated by the NumWord transformation.Source: §1 · p. 2
  • GPT-3.5 outperforms existing LLMs in most NLU tasks in zero- and few-shot settings, especially reading comprehension, natural language inference, and semantic matching.Compared with zero-shot prompting, one-shot performs better on most tasks; additional examples can benefit sequence tagging.Source: §1 · pp. 2–3
  • The authors suggest that training stages, task-label alignment, and instruction design matter for comprehension and application performance, while noting that the proposed stage-level explanation is speculative.On IMDB, GPT-3.5 produces many “neutral” responses outside the binary application label space; the authors also call for research on instruction robustness.Source: §1 · pp. 2–3

2 Background and Related Work

Prior work has emphasized LLM task performance, while robustness has received less attention despite its relevance to open-world complexity and trustworthy AI. The study uses TextFlint to evaluate GPT-3.5 robustness.

2.1 Large Language Models

  • Prior work has largely examined LLM performance across tasks, with comparatively little attention to robustness, a dimension relevant to handling open-world complexity and trustworthy AI.The paper conducts a comprehensive experimental analysis of GPT-3.5 robustness.Source: §2.1 · p. 4

2.2 TextFlint

  • TextFlint provides text transformations and perturbation types for robustness evaluation, and the study uses 66 of its transformations.The toolkit includes universal and task-specific transformations, adversarial attacks, subpopulations, and combinations.Source: §2.2 · p. 4

3 Experiment Setup

The experiments cover nine NLU task categories and compare GPT-3.5 with earlier GPT, supervised, and zero-/few-shot models. They test zero- and few-shot settings and select three prompts per dataset for evaluation.

3.2 Task and Datasets

  • The evaluation covers 21 datasets across nine NLU categories, applying 66 TextFlint transformations to assess performance and robustness.The categories include reading comprehension, relation extraction, inference, semantic matching, sentiment analysis, named entity recognition, part-of-speech tagging, and WSC.Source: §3.2 · p. 5

3.3 Model Selection

  • The study compares GPT-3.5 models with text-davinci-001, supervised models, and zero-/few-shot models.text-davinci-002 is selected from the GPT-3.5 series for comparison with GPT-3 text-davinci-001.Source: §3.3 · p. 5

3.4 Zero/few-shot Scene Setting

  • Models are evaluated in zero-shot and 1-, 3-, 6-, and 9-shot settings, but 6- and 9-shot tests use only 1,000 original-dataset samples.The 1- and 3-shot evaluations cover the full dataset with task-specific transformations.Source: §3.4 · p. 5

3.5 Prompt Selection

  • The researchers select the three best-performing prompts per dataset and expand them to five shot levels; few-shot results are means and standard deviations across those prompts.Prompts are collected or manually designed for particular tasks.Source: §3.5 · p. 5

4 Experiments

The experiments compare zero- and few-shot GPT models with fine-tuned models on machine reading comprehension and relation extraction. Results vary by task: few-shot models are competitive in reading comprehension, while relation extraction remains a substantial challenge for GPT models.

4.1 Machine Reading Comprehension

  • Few-shot models are comparable to fine-tuned models in reading-comprehension performance and robustness, with GPT models more robust on some SQuAD1.1 transformations.The number of prompt examples has inconsistent effects: robustness remains relatively unchanged, while 1- and 3-shot micro-F1 and EM improve over zero-shot.Source: Table 2 · p. 6, Table 3 · p. 6, §4.1 · p. 7

4 Experiments

  • Across the experiments, GPT-series models are evaluated in zero- and few-shot settings against fine-tuned models on machine reading comprehension and relation extraction.The experiments also compare robustness under different tests, using SQuAD1.1 and SQuAD2.0 for reading comprehension and Tacred for relation extraction.Source: Table 2 · p. 6, Table 3 · p. 6, §4.1 · p. 7

4.2 Relation Extraction

  • On Tacred, fine-tuned models reach 70% micro-F1 with large training data, while the best GPT-series performance is only about 20%.GPT models nevertheless perform accurately on the specific person-and-age relation, where the dataset’s fixed syntax and conventional transformed expressions provide prediction hints.Source: Table 6 · p. 8, §4.2 · p. 8

4.3 Sentence Pair Relationship

The section evaluates natural language inference and semantic matching with zero- and few-shot GPT models under text transformations. Robustness differs across transformations and tasks, with prompt examples associated with improved robustness on MRPC semantic matching.

4.3.1 Natural Language Inference

  • For MNLI-m and MNLI-mm, AddSent robustness is slightly lower for zero-, one-, and three-shot models than for fine-tuned models, while SwapAnt robustness is significantly better.NumWord robustness is comparable between GPT and fine-tuned models.Source: Table 7 · p. 9, Table 8 · p. 9, §4.3.1 · p. 10

4.3.2 Semantic Matching

  • On MRPC and QQP, GPT-3.5 shows greater robustness than fine-tuned models on NumWord and SwapAnt transformations, especially on MRPC.Fine-tuned MRPC accuracy on transformed data is often close to zero, whereas the GPT series does not show that pattern.Source: Table 10 · p. 10, §4.3.2 · p. 11, Table 11 · p. 11
  • On MRPC, one- and three-shot robustness is better than zero-shot, but there is little difference between one and three examples.For NumWord-transformed MRPC data, one- and three-shot accuracy reaches 100%.Source: Table 10 · p. 10, §4.3.2 · p. 11

4.4 Sentiment Analysis

The sentiment-analysis and sequence-tagging experiments show uneven robustness and performance across tasks and prompt settings. Few-shot learning benefits aspect-based sentiment analysis, while sentiment classification and tagging expose distinct limitations and sensitivities.

4.4.1 Aspect-based Sentiment Analysis

  • For aspect-based sentiment analysis, zero-, one-, and three-shot models perform comparably to fine-tuned models, and one- and three-shot robustness is comparable to or better than fine-tuned robustness.Results are consistent across SemEval2014-Laptop and SemEval2014-Restaurant, with little performance variation between prompts.Source: Table 12 · p. 12, Table 13 · p. 12, §4.4.1 · p. 13

4.4.2 Sentiment Classification

  • On IMDB sentiment classification, neutral answers increase as prompt examples increase, contributing to declining three-shot performance.Shorter prompt examples perform better than longer ones on this dataset.Source: §4.4.2 · p. 13, Table 14 · p. 13

4.5.1 Named Entity Recognition

  • On ACE2005, GPT-series best NER F1 scores hover around 50, compared with nearly 90 for fine-tuned models, and robustness varies across transformations.Adding prompt examples can improve GPT-series performance, but transformation effects on robustness vary by dataset and transformation.Source: Table 15 · p. 14, §4.5.1 · p. 15

4.5.2 Part-of-speech Tagging

  • On WSJ part-of-speech tagging, GPT-series accuracy remains below fine-tuned models, though the gap is smaller than in NER, and three-shot performance improves over zero-shot.Text-davinci-001 fails to complete the task under the prompt, while text-davinci-002 and -003 show robustness roughly equal to fine-tuned models.Source: Table 18 · p. 16, §4.5.2 · p. 17

4.6 The Winograd Schema Challenge

On WSC273, text-davinci-003 performs near or above BERT-large but remains significantly behind state-of-the-art results. Greater GPT-model capability does not ensure greater robustness across transformations or every dataset variant.

4.6 The Winograd Schema Challenge

  • On WSC273, text-davinci-003 is comparable to BERT-large but remains significantly different from state-of-the-art results.Its improved capability is accompanied by reduced robustness on some text transformations.Source: §4.6 · p. 17, Table 19 · p. 17
  • The best-performing model varies across WSC dataset variants: text-davinci-003 leads on the original dataset, while text-davinci-002 and text-davinci-001 lead on name- and gender-related datasets, respectively.Source: §4.6 · pp. 17–18
  • The authors state that the enhanced model capability may have lost some careful judgment and needs further investigation.Source: §4.6 · p. 18

4.7 Further Studies

Few-shot prompting improves some partial sequence-tagging results, though scores remain low; text-davinci-003 also outperforms earlier LLMs on most compared tasks at 3-shot. It exceeds their zero-shot best on some tasks, including NLI and sentiment analysis.

4.7.1 Partial Sequence Tagging Task

  • Some partial sequence-tagging results are unavailable because the model did not complete the task as instructed, preventing analysis.Source: Table 20 · p. 18
  • Partial sequence-tagging scores remain low, but adding prompt examples significantly improves results on most datasets; text-davinci-003 performs best overall in its 3-shot setting.The evaluation uses accuracy for Daily547 and micro-F1 for the other datasets.Source: §4.7.1 · p. 18, Table 20 · p. 18

4.7.2 Comparison of text-davinci-003 and other LLMs

  • At 3-shot, text-davinci-003 outperforms the original LLMs’ best few-shot performance on most compared tasks, and exceeds their zero-shot best on some tasks.The latter includes NLI and sentiment analysis.Source: Table 21 · p. 18, §4.7.2 · p. 19

5 Conclusion

The study evaluates GPT-3.5 robustness across 21 datasets and 9 NLU tasks. It finds task-dependent advantages over fine-tuned models alongside substantial degradation and several specific robustness challenges.

  • Across 21 datasets and 9 NLU tasks, GPT-3.5 performs better than existing fine-tuned models on some tasks but still shows significant robustness degradation.Source: §5 · p. 19
  • The study identifies robustness instability, prompt sensitivity, number sensitivity, and sensitivity to task labels and types as specific challenges.Source: §5 · p. 19
  • The authors call for further study of instruction robustness, more consistent prompting between instruction tuning and deployment, and better prompts for complex tasks such as NER and RE.They note that instruction-construction complexity may contribute to lower performance on NER and RE.Source: §5 · p. 19

A Details of Performance

This appendix identifies the contents of the performance table and its accompanying visualization.

B Prompts

The appendix supplies dataset-specific prompt templates. It provides 0- and 1-shot prompts, with 3-, 6-, and 9-shot variants adding further in-context examples.

Related papers

Loading 2303.00293v1…