Evaluation Metrics in the Era of GPT-4: Reliably Evaluating Large Language Models on Sequence to Sequence Tasks

1 Introduction

• The study evaluates open- and closed-source LLMs on summarisation, simplification, and GEC using automatic, human, and GPT-4 evaluations.The tasks are reframed as text generation without task-specific fine-tuning.Source: §1 · p. 1
• The evaluation addresses prior work often limited to a single task, model, dataset, or automatic evaluation.The paper describes its evaluation as multidataset, multi-model, and multi-task.Source: §1 · p. 1
• Traditional reference-based metrics do not reliably predict human judgments, and human reviewers rate gold references below model outputs.The authors say it is unclear whether this misalignment reflects metric limitations, poor reference quality, or both.Source: §1 · pp. 1–2
• GPT-4 shows reasonable alignment with human judgment when ranking different models across most tasks and metrics, although alignment varies, with lower alignment on some metrics than others.Source: §1 · p. 2

2.1 Datasets

• The study uses 3,000 Newsela test samples for simplification, 3,000 CNN/DailyMail test samples for summarisation, and 4,384 BEA-2019 development samples for GEC.The Newsela samples were randomly selected after removing redundancy; the BEA-2019 test set was not used because it was not fully disclosed.Source: §2.1 · p. 2, §2.2 · p. 2

2.2 Models

• All experiments are zero-shot and use no additional fine-tuning or in-context learning.Some models were already fine-tuned to follow instructions across NLP tasks.Source: §2.2 · p. 2
• Experiments use Flan-T5, T0pp, OPT-IML, Flan-UL2, GPT-3, InstructGPT, and ChatGPT.The paper describes these as a range of open-source LLMs and OpenAI commercial models.Source: §2.2 · p. 2
• Prompt engineering and hyper-parameter tuning details are provided in Appendix B.The study's prompt investigation was limited, as discussed in the paper's limitations.Source: §2.2 · p. 2

3 Evaluation Metrics

• Model outputs are assessed with automatic metrics, human evaluation, and GPT-4 as a reviewer.GPT-4 is included to assess whether it can be reliably deployed in place of human reviewers.Source: §3 · p. 2, §3.3 · p. 3

3.1 Automatic Evaluation

• Automatic evaluation uses SARI for simplification, geometric-mean ROUGE-{1, 2, L} F1 for summarisation, and ERRANT F0.5 for GEC.These are the most widely adopted reference-based metrics the paper selected for each task.Source: §3.1 · p. 2

3.2 Human Evaluation

• Three reviewers assess 100 randomly selected samples per dataset, rating both model outputs and the gold reference.Reviewers use task-specific criteria on five-point Likert scales; the GEC criteria include Semantics, Grammaticality, and Over-correction.Source: §3.2 · p. 2

3.3 GPT-4 as a Reviewer

• GPT-4 reviews the same 100 samples per task as the human reviewers, using the criteria and assessment scales included in its prompt.The prompt also includes the input text for each sample.Source: §3.3 · p. 3

4.1 Automatic Evaluation Results

• T0pp significantly outperforms GPT-3 and ChatGPT on summarisation; Flan-T5 and InstructGPT lead simplification; ChatGPT and OPT-IML perform similarly on GEC.The reported differences versus GPT-3 on simplification and GEC are significant at p < 0.001.Source: §4.1 · p. 3
• The same prompt appears best across models and temperature settings for each task, with one exception.The authors interpret this pattern as suggesting that prompt quality is almost model-invariant.Source: §4.1 · p. 3

4.2 Human and GPT-4 Evaluation Results

• Human reviewers show generally strong agreement, with an average Krippendorff α of 0.88 across metrics and a lowest value of 0.62.The study converts human and GPT-4 scores to rankings before comparing them.Source: §4.2 · p. 3
• On summarisation, human reviewers favor ChatGPT on several criteria despite its worse ROUGE score, and all rate gold summaries worst on every criterion.Most reviewers rate ChatGPT best for Relevance and Fluency; all rate it best for Coherence and Consistency.Source: §4.2 · p. 3
• On simplification, reviewers rank ChatGPT best for Fluency and Simplicity but rate it poorly for Semantics; Flan-T5 leads Semantics after often copying inputs unchanged.All reviewers score the gold reference worst.Source: §4.2 · p. 4
• On GEC, reviewers rank ChatGPT best for Grammaticality but worst or second worst for Semantics and Overcorrection, where OPT-IML leads.The gold reference is mostly second worst on most metrics and for most reviewers.Source: §4.2 · p. 4
• GPT-4 rankings correlate well with human judgments on summarisation and simplification, but disagree more on GEC.Including GPT-4, average Krippendorff α is 0.70 across metrics, with a lowest value of 0.34.Source: §4.2 · p. 4

• The study examined zero-shot model performance across open-source datasets, combining automatic, human, and GPT-4 model-to-model evaluation.Source: §5 · p. 4
• Reference-based metrics are systematically misaligned with human evaluation across generative tasks, and the authors cannot determine whether metric limitations or poor references explain the mismatch.Source: §5 · pp. 4–5
• The best open-source LLMs consistently outperformed the reference outputs in human evaluation.Source: §5 · p. 5
• GPT-4’s judgments correlated strongly with human judgments for summarisation and simplification, and moderately for GEC.Source: §5 · p. 5
• Future work includes larger human evaluations, broader model settings, and multiple automatic metrics, including reference-less metrics.The authors note that metric-based selection of cases for human evaluation can introduce bias.Source: §5 · p. 5

Limitations

• Only 2 to 5 prompts per task were tried, leaving focused prompt-engineering investigation for future work.The authors say it could potentially improve results, but treat it as a stand-alone exploration topic.Source: Limitations · p. 5
• The experiments omitted in-context learning and chain-of-thought prompting, so model outputs may have room to improve.The authors say gold-reference quality would remain unchanged until new datasets become available.Source: Limitations · p. 5
• Automatic metrics selected settings for human evaluation, potentially biasing which outputs were reviewed; one gold reference can also mis-score equally valid outputs.The authors illustrate that two good summaries may share few bi-grams and receive a poor ROUGE score.Source: Limitations · p. 5
• Some evaluated models may have encountered the open-source test datasets during pretraining, potentially affecting results.The authors say using new unreleased datasets would have been preferable, but creating high-quality datasets from scratch was costly and time-consuming.Source: Limitations · p. 5
• GPT-4 may favor ChatGPT outputs because both belong to the OpenAI model family, but the study could neither validate nor refute that bias.Human reviewers also rated ChatGPT best across most metrics.Source: Limitations · p. 5
• Three reviewers annotated 100 samples per dataset, which the authors say may limit the strength and generalisability of the findings.They note that annotators consistently judged the gold standard worse than the best model outputs across tasks.Source: Limitations · pp. 5–6

Ethics Statement

• The study used LLMs and open-source datasets without fine-tuning, while some evaluated OpenAI models were not open-source.The authors state that the research explored model limitations and performance and did not increase risks caused by the models or datasets.Source: Ethics Statement · p. 6
• Three human annotators participated in a 16-hour study and were paid £13.20 per hour, excluding Prolific fees.Source: Ethics Statement · p. 6
• The researchers did not collect personal information beyond Prolific-provided demographic data and did not use those data as screening criteria.Annotators received a detailed study description before agreeing to participate.Source: Ethics Statement · p. 6

A Newsela Dataset Processing

• When either the source or destination was duplicated, the authors merged the samples and combined the distinct destinations in their original order.For duplicate sources with different destinations, they kept the source once and merged the destinations; they applied the converse procedure when destinations were duplicated.Source: A Newsela Dataset Processing · p. 8
• The example combines two different destinations for the repeated Ron Bee source into one destination text.The combined text retains the California professor description and adds that very few people join the military now.Source: A Newsela Dataset Processing · p. 8

B Implementation Details

• Full experiments used the most promising settings selected through preliminary tests on 100 samples per dataset, with temperature choices varying by task.Lower temperatures performed best for summarisation and simplification, while some GEC models benefited from higher values.Source: B Implementation Details · p. 8, §B.2 · p. 9

B.1 Prompt Engineering

• The same prompt generally worked best across models and temperatures, with only extremely limited exceptions.The appendix identifies the prompt used for evaluation as prompt (a) for each task.Source: §B.1 · p. 8

1. Text summarisation

• Summarisation prompts varied in their lead-in and requested summary framing, including a “very short summary” variant.Other variants asked for a summary of the text or framed it as the main story.Source: §1 · pp. 8–9

2. Text simplification

• Simplification prompts ranged from direct simplification requests to asking for an explanation suitable for a five-year-old.The listed alternatives also varied how the input and requested output were framed.Source: §2 · p. 9

3. Grammatical Error Correction

• GEC prompts asked for corrected sentences, with one explicitly instructing models to fix grammatical and spelling errors and preserve unchanged sentences when error-free.An alternative asked for standard English corrections.Source: §3 · p. 9

B.2 Hyperparameter Tuning

• The best overall temperature was 0.0, though some GEC model-prompt combinations did better at 0.2 or 0.5.For HuggingFace models, 0.01 substituted for 0.0 due to implementation requirements; other hyperparameters remained at defaults.Source: §B.2 · p. 9

B.3 Tokenization and Truncation

• Because CNN/DailyMail inputs often exceeded model limits, truncation thresholds were tailored to each model’s input-output token constraints.For example, 80.6% of CNN/DailyMail samples exceeded 482 tokens and 9.8% exceeded 1,506 tokens; GPT-3 inputs were not truncated, while other models used specified limits.Source: §B.3 · p. 9

C Human Evaluation Criteria for GEC

• GEC reviewers separately assessed semantic preservation, remaining grammatical errors, and unnecessary edits, since over-correction can occur without meaning change.Semantic preservation used a five-point scale, grammaticality counted remaining errors, and over-correction was rated by severity.Source: C Human Evaluation Criteria for GEC · p. 10

D Detailed Automatic Evaluation Results

• Table 3 presents detailed automatic evaluation results for text summarisation.The reported experiments vary model, temperature, and prompt.Source: Table 3 · p. 10, D Detailed Automatic Evaluation Results · p. 10
• Table 4 presents detailed automatic evaluation results for text simplification.The reported experiments vary model, temperature, and prompt.Source: D Detailed Automatic Evaluation Results · p. 10, Table 4 · p. 11
• Table 5 presents detailed automatic evaluation results for GEC.The reported experiments vary model, temperature, and prompt.Source: D Detailed Automatic Evaluation Results · p. 10, Table 5 · p. 11

E Detailed Human and GPT-4 Evaluation Results

• Table 6 reports average scores and standard deviations from human annotators and GPT-4 across models, tasks, and metrics.Results cover 100 analysed samples per task.Source: Table 6 · p. 12, E Detailed Human and GPT-4 Evaluation… · p. 12
• Table 7 reports model rankings by annotator and task-metric Krippendorff α agreement coefficients, including GPT-4 comparisons.Rankings use 1 for best and 4 for worst; agreement coefficients are aggregated across models.Source: Table 7 · p. 13