LLMs as Narcissistic Evaluators: When Ego Inflates Evaluation Scores

1 Introduction

• Evaluating generated text is challenging because different forms can convey the same meaning, while expert human evaluation is costly and time-consuming.Automatic metrics have emerged as practical alternatives.Source: §1 · p. 1
• The study examines six model families and their corresponding evaluators or conditional generative probability, varying model size and fine-tuning settings.The models serve as both generators and evaluators.Source: §1 · pp. 1–2
• Experiments use CNN/Daily Mail and XSUM summaries in reference-based evaluation with gold summaries and reference-free evaluation against source documents.The passage describes the former as common in supervised summarization and the latter in unsupervised summarization and factuality assessment.Source: §1 · p. 2
• Self-bias grows more pronounced when generator and evaluator match in fine-tuning configuration and model size.The authors also report inflated scores beyond self-bias, including a preference for longer summaries by certain evaluators.Source: §1 · p. 2
• Inflated scores are particularly noticeable in reference-free evaluation, a popular approach for assessing factual correctness.The passage cites Koh et al. (2022) in connection with the popularity of this evaluation approach.Source: §1 · p. 2

2 Related Work

• BERTScore, MoverScore, BARTScore, and BLEURT use pretrained language models for semantic-level assessment beyond direct word overlap (Zhang et al., 2020; Zhao et al., 2019; Yuan et al., 2021; Sellam et al., 2020).Source: §2 · p. 2
• Reference-free generative metrics such as BARTScore and GPTScore evaluate text against source documents, a setting used for factual consistency assessment.The passage notes that BARTScore in reference-free mode had the highest overall correlation with human factual-consistency scores in long-document abstractive summarization.Source: §2 · p. 2
• Automatic metrics also face robustness concerns, including spurious correlations, noise, out-of-domain text, long documents, and potential preference for extractive summaries (Sai et al., 2021; Amplayo et al., 2022).Source: §2 · p. 2
• Traditional ROUGE and BLEU are criticized for missing content quality and syntactic errors, while embedding metrics remain inadequate for measuring shared information between summaries.Traditional metrics may score fluent but meaningless or unfaithful outputs highly when they reuse words (Gehrmann et al., 2021).Source: §2 · pp. 2–3
• Prior work reports that reference-free metrics can favor outputs from models more similar to themselves, while systematic examination across models and settings remained open.The authors frame their study around that gap, covering both reference-based and reference-free evaluation.Source: §2 · p. 3

3 Methodology

• Quantitative comparisons vary evaluator and generator model size and fine-tuning configuration; qualitative analysis checks whether self-bias scores track human judgments.The qualitative check uses summaries with human evaluations.Source: §3 · p. 3

3.1 Evaluators

• BART-, T5-, GPT-based evaluators and Cohere use conditional probabilities, scoring a hypothesis with an evaluation aspect and supplementary source or reference text.The study includes GPTScore variants and Cohere; the latter follows the scoring approach used by BARTScore, T5Score, and GPTScore.Source: §3.1 · pp. 3–4

3.2 Generation Models

• Generation models span BART, T5, GPT-2, GPT-3, FLAN-T5, and Cohere, with available size variants and pretrained, CNN-fine-tuned, or XSUM-fine-tuned settings.Models serve as both summary generators and underlying LM evaluators; publicly available checkpoints are used, with stated exceptions for GPT-3-Curie and Cohere.Source: §3.2 · p. 4

3.3 Datasets

• The study uses CNN/DailyMail and XSUM for quantitative comparisons, and human-rated SummEval and RoSE summaries for qualitative analysis.It samples 500 documents per dataset for quantitative comparisons; SummEval covers four human-rated aspects, while RoSE uses ACU annotations based on facts in reference summaries.Source: §3.3 · p. 4

3.4 Quantitative Comparisons

• Twenty evaluators compare summaries with source documents in reference-free evaluation or with reference summaries in reference-based evaluation; their scores are uniformly normalized for comparability.The analysis also examines summary length and its correlation with evaluator scores.Source: §3.4 · p. 5

3.5 Qualitative Analysis

• Qualitative analysis uses Spearman and Kendall correlations with human judgments, pairing SummEval aspects with reference-free evaluation and RoSE ACU annotations with reference-based evaluation.SummEval aspects are coherence, consistency, fluency, and relevance.Source: §3.5 · p. 5

4.1 Quantitative Comparisons: Assessing Self-Bias in LM-Evaluators Towards Their Own Output

• Diagonal heatmap cells show evaluators favoring their corresponding generators in both settings, with stronger self-bias in reference-free evaluation.The CNN/DM heatmaps average evaluator scores over 500 documents.Source: §4.1 · p. 5, Figure 2 · p. 6, Figure 3 · p. 7
• T5-based generators receive higher scores from T5Score variants, while BART-based generators show a parallel preference from BARTScore variants.The T5 preference appears as a concentrated dark rectangle in the heatmap.Source: §4.1 · pp. 5–6
• Evaluators tend to rank generators trained on the same dataset more highly, including T5-XSUM evaluators preferring BART-XSUM over T5-vanilla on CNN Daily.The same pattern also appears on summaries generated from XSUM documents.Source: §4.1 · p. 6

4.2 Bias towards Longer Summaries

• BARTScore and GPTScore variants generally assign higher scores to longer summaries, except for XSUM-finetuned evaluators; this length correlation is less obvious in reference-based evaluation.Cohere and XSUM-finetuned evaluators show a preference for shorter summaries.Source: §4.2 · p. 6, Figure 5 · p. 8

4.3 Qualitative Analysis: Correlation of Self-Bias with Human Evaluation

• On SummEval and RoSE, T5-base receives higher scores from T5 evaluators, while BART-based models receive higher scores from BARTScore and GPTScore evaluators.SummEval is evaluated reference-free and RoSE reference-based, matching the aspects emphasized by each benchmark.Source: §4.3 · pp. 7–8, Figure 4 · p. 7
• None of the evaluators strongly correlates with human annotations; samples of 100 summaries per benchmark and missing generators prevent a conclusive correlation assessment.The authors note that inflated scores for an evaluator's own generator may contribute to the low correlation.Source: §4.3 · p. 8

• In SummEval, T5 is ranked higher only by certain T5Score variants in the reference-based setting.Source: §5 · p. 8
• Models fine-tuned on XSUM are not suitable for direct evaluator use because of bias toward shorter summaries, except for evaluating summaries aligned with XSUM-style content.Source: §5 · p. 8
• Self-bias appears across all assessed evaluators, so the authors recommend avoiding the same underlying model for generation and evaluation.Source: §5 · p. 8
• Contemporary evaluation metrics may favor longer summaries, a bias to consider when interpreting and applying their scores.Source: §5 · p. 8
• Limited human evaluations prevent definitive conclusions about which examined generative evaluator is best, while the work points toward developing more resilient and unbiased metrics.Source: §5 · p. 8
• The authors identify a new type of bias in generative evaluators and encourage future research toward fairer evaluation metrics.Source: §5 · p. 8
• In the reference-free XSUM heatmap, the darkest diagonal cells indicate evaluator preference for their respective models; scores are normalized from 0 to 1 and brackets give average summary lengths in words.Source: Figure 6 · p. 9

• The human evaluations are constrained by limited sample sizes from SummEval and RoSE, motivating larger-scale human evaluation.Source: Limitations · p. 9
• The experiments focus on summarization; extending the evaluation to a broader range of generation tasks would be beneficial.Source: Limitations · p. 9
• The authors say incorporating additional generation models and evaluators would further enrich the experiment.Source: Limitations · p. 10

• The authors report no ethical concerns and note that the study's data and supplementary materials are open-sourced and widely employed in existing works.Source: Ethics Statement · p. 10

A.1 Generator

• Generator model details, including checkpoints and prompt settings, are listed in Table 4.Source: §A.1 · p. 12

A.2 Evaluator

• Evaluator model details are listed in Table 5 for reference-free settings and Table 6 for reference-based settings.Source: §A.2 · p. 12

B.1 Reference-free Setting

• Figure 6 presents XSUM evaluation results in the reference-free setting.Source: §B.1 · p. 12
• Figure 4 presents RoSE and SummEval evaluation scores in the reference-free setting.Source: §B.1 · p. 12
• Table 3 presents SummEval Spearman and Kendall correlations for reference-free meta-evaluation.Source: §B.1 · p. 12

B.2 Reference-based Setting

• Figure 3 presents a heatmap of CNN/DM evaluation results in the reference-based setting.Source: §B.2 · p. 12
• Table 4 lists generation models or checkpoints with their prompt configurations.Source: Table 4 · p. 13
• Table 5 lists models or checkpoints and prompt configurations used for reference-free evaluation.Source: Table 5 · p. 13
• Figure 4 illustrates RoSE and SummEval evaluation scores in the reference-based setting.Source: §B.2 · p. 13
• Table 2 presents RoSE benchmark results.Source: §B.2 · p. 13
• The passage identifies Spearman and Kendall correlations for reference-based meta-evaluation but does not complete the benchmark identification.Source: §B.2 · p. 13
• Table 6 lists models or checkpoints and prompt configurations used for reference-based evaluation.Source: Table 6 · p. 14