Image Captioning Evaluation in the Age of Multimodal LLMs: Challenges and Future Perspectives

1 Introduction

Image captioning, the task of generating natural language descriptions of visual content [Stefanini et al., 2022], is a fundamental component of modern Artificial Intelligence systems such as Multimodal Large Language Models (MLLMs) and autonomous agents. These models, indeed, need to understand and interact with visual data effectively and be capable of generating high-quality descriptions. As MLLMs continue to advance, the development of robust and reliable evaluation metrics for image descriptions becomes essential.

The history of image description has been marked by continuous innovation. Early captioning models evolved from template-based and recurrent neural networks [Vinyals et al., 2015] to Transformer-based architectures [Cornia et al., 2020; Pan et al., 2020]. Typically, an image captioning model consists of a visual encoder and a language model. Visual encoders range from CNNs [Anderson et al., 2018] to Vision Transformers [Dosovitskiy et al., 2021], while language models have transitioned from LSTMs to Transformers [Vaswani et al., 2017]. More recently, MLLMs, such as LLaVA [Liu et al., 2024], IDEFICS [Laurenc¸on et al., 2023], and Llama 3.2 Vision [Dubey et al., 2024], have integrated these capabilities, transforming how captions are generated and evaluated.

Training strategies have also evolved, incorporating supervised learning integrating both cross-entropy loss and reinforcement learning using self-critical sequence training [Rennie et al., 2017]. Several benchmark datasets have facilitated the development of image captioning models, including COCO [Lin et al., 2014], Conceptual Captions [Sharma et al., 2018], and nocaps [Agrawal et al., 2019], each providing varying levels of diversity and complexity.

As captioning models improve, the role of evaluation metrics becomes increasingly critical. A good metric must assess captions across multiple dimensions to ensure alignment with human expectations both linguistically and semantically. A strong captioning metric should reward captions that are fluent, accurate, and faithful to the image content, while also ensuring they focus on the most salient aspects and strike a balance between detail and conciseness. Furthermore, it should penalize hallucinated or misleading information.

Early metrics such as BLEU [Papineni et al., 2002], ROUGE [Lin, 2004], and METEOR [Banerjee and Lavie, 2005] were borrowed from machine translation and focused on n-gram overlap with reference captions. Later, specialized metrics like CIDEr [Vedantam et al., 2015] and SPICE [Anderson et al., 2016] aimed to improve correlation with human judgment by incorporating consensus-based scoring and scene graph analysis, respectively. While these metrics introduced refinements, they remained constrained by their reliance on reference captions and linguistic overlap.

Recent advances have shifted towards learnable metrics leveraging vision-and-language pre-trained models. CLIPS [Hessel et al., 2021] introduces a contrastive approach to measure image-caption alignment based on CLIP embeddings. Fine-grained metrics like BRIDGE [Sarto et al., 2024a] and HiFi-S [Yao et al., 2024] capture localized semantics and hierarchical structures. LLM-based metrics such as CLAIR [Chan et al., 2023] and FLEUR [Lee et al., 2024] integrate reasoning and explanation into caption evaluation.

With the rise of MLLMs as image description generators, the nature of captions has changed, as these models generate captions that differ in length and specificity. Conventional evaluation metrics may no longer reliably assess caption quality in this new paradigm, posing new challenges for the task. Following this insight, this survey provides a comprehensive analysis of existing captioning metrics, from traditional rulebased approaches to modern learnable methods and LLMdriven techniques. We explore their evolution, strengths, and limitations while emphasizing the growing need for robust evaluation frameworks that can adapt to the stylistic diversity introduced by MLLMs. Additionally, we present an in-depth experimental study assessing metrics across multiple dimensions, including correlation with human judgment, ranking accuracy, and sensitivity to hallucinations. Our analysis highlights the performance gaps of conventional metrics and showcases promising directions for future research.

2 The Evolution of CaptioningMetrics

In this section, we present the evolution of captioning metrics, outlining their development from traditional non-learningbased evaluation strategies to recent advancements incorporating LLMs. Fig. 1 presents a taxonomy that highlights the key characteristics of the most commonly used metrics, including their reliance on reference captions or image input.

2.1 Rule-basedMetrics

Many widely used evaluation metrics for image captioning originate from NLP tasks and rely heavily on n-gram matching techniques. Popular metrics in this domain include BLEU [Papineni et al., 2002], METEOR [Banerjee and Lavie, 2005], ROUGE [Lin, 2004], CIDEr [Vedantam et al., 2015], and SPICE [Anderson et al., 2016]. Despite their extensive use in image captioning, these traditional metrics exhibit several limitations. Notably, they do not consider the input image during evaluation and often demonstrate weak correlations with human judgment. Also, their reliance on reference captions restricts their ability to comprehensively evaluate generated captions, as their effectiveness is inherently influenced by the style and quality of the reference dataset.

BLEU [Papineni et al., 2002] and METEOR [Banerjee and Lavie, 2005] were originally designed for machine translation tasks. BLEU evaluates n-gram precision and introduces the concept of modified unigram precision, which adjusts word frequency counts by considering the maximum occurrence of each word across all reference translations. By design, it penalizes overly long translations through precision-based scoring and mitigates excessively short ones with a brevity penalty, promoting balance in word choice, order, and length. However, BLEU lacks a direct measure of recall, which quantifies the proportion of matched n-grams in the reference. Recall is particularly crucial for evaluating translation quality, as it reflects the extent to which a candidate text captures the content of the reference. Introduced to mitigate this limitation, METEOR prioritizes recall by considering unigram matches between candidate and reference sentences and incorporating exact matches, stemmed forms, and semantic equivalences. This enables METEOR to provide a more detailed and holistic assessment of translation quality. Similarly, ROUGE [Lin, 2004], initially developed for summarization tasks, has been adapted for evaluating visual descriptions and emphasizes n-gram recall by analyzing the overlap of n-grams between candidate and reference texts.

Recognizing the limitations of these general-purpose metrics, CIDEr [Vedantam et al., 2015] and SPICE [Anderson et al., 2016] were designed for the captioning task. CIDEr employs a TF-IDF weighting scheme to reweight the significance of different n-grams, capturing both precision and recall while accounting for their contextual importance. SPICE takes a structured approach by converting captions into scene graphs and measuring their similarity, aligning the evaluation with the semantic and structural elements of the captions.

2.2 Learnable Metrics

All the aforementioned metrics primarily rely on textual-level comparisons and struggle to adequately account for synonym matches from a linguistic perspective. Moreover, they assume that human-written reference captions perfectly reflect the content of the image, which may not always be the case. To overcome these limitations, recent metrics leverage pretrained models to compare textual-only and visual-textual content, using both supervised and unsupervised approaches.

UnsupervisedMetrics. BERT-S [Zhang et al., 2020], for instance, utilizes learned embeddings from the pre-trained language model BERT [Devlin et al., 2018] to more effectively measure semantic similarities between candidate and reference captions. Building on this, BERT-S++ [Yi et al., 2020] enhances the approach by incorporating the variance across multiple reference captions, improving robustness. However, these metrics still do not integrate the visual modality, which can limit their ability to fully evaluate the performance of captioning models. To overcome this, metrics that explicitly include the image as input have been introduced. These methods leverage pre-trained vision-and-language models, such as UNITER [Chen et al., 2020], ViLBERT [Lu et al., 2019], and CLIP [Radford et al., 2021], to evaluate captions in a way that aligns more closely with both textual and visual content.

Among these metrics, UMIC [Lee et al., 2021] stands out as it effectively addresses a key limitation of BERT-S: the difficulty of evaluating image captions in the absence of sufficient reference captions. To overcome this, it introduces a reference-free approach built on a fine-tuned version of a pretrained multimodal Transformer (i.e., UNITER). By leveraging contrastive learning, the model is trained to differentiate between ground-truth captions and diverse synthetic negative samples. These negative samples are meticulously designed to cover various undesirable captioning cases, such as those that are grammatically incorrect, irrelevant to the image, or relevant but containing incorrect keywords. This approach ensures a more robust evaluation of image captioning models without relying on reference captions.

TIGEr [Jiang et al., 2019], on the other hand, focuses on the integration of the image as one of the inputs for computing an evaluation score. Building upon a text-image grounding model [Lee et al., 2018], it allows evaluating caption quality not only based on how well a caption represents image content but also on how well machine-generated captions match human-generated ones. Following this line and inspired by BERT-S, ViLBERT-S [Lee et al., 2020] computes cosine similarity between token embeddings for reference and candidate sentences. However, different from BERT-S, the token embedding is computed with the consideration of image contexts. On a different line, FAIEr [Wang et al., 2021] employs the scene graph as a bridge between the image and textual modality. Specifically, it fuses scene graphs of the image and references as a union scene graph and compares it with the scene graph of generated captions.

Leveraging the success of employing pre-trained visionand-language models, numerous metrics started exploiting the CLIP extensive pre-training to compute image captioning evaluation scores. Among these, CLIP-S [Hessel et al., 2021] was the first metric to utilize a modified cosine similarity between image and candidate caption representations derived from the visual and textual encoders of CLIP. In a similar way, MID [Kim et al., 2022] leverages CLIP-based visualtextual features to compute negative Gaussian cross-mutual information, yielding a more effective evaluation metric.

While these advancements highlight the potential of contrastive-based embedding spaces for caption evaluation, large-scale models pre-trained on web-sourced data have inherent limitations. In particular, CLIP, which is trained on web-collected image-caption pairs, may be suboptimal for image evaluation tasks, as such annotations often lack the richness, grammatical correctness, and detail required to assess generated, typically longer captions. To address this, methods like PAC-S [Sarto et al., 2023] and its improved version PAC-S++ [Sarto et al., 2024b] aim to refine the CLIP embedding space by incorporating a fine-tuning phase. This process incorporates curated image-caption pairs along with synthetically generated positive examples, achieving improved performance in image captioning evaluation.

Supervised Metrics. All the aforementioned metrics leverage the CLIP embedding space, either pre-trained or finetuned on image-caption pairs. However, CLIP was not originally designed specifically for computing evaluation scores, which can limit its effectiveness. To address this challenge, a less explored research direction focuses on refining the CLIP embedding space through supervised training approaches. For instance, the Polos metric [Wada et al., 2024] is based on Polaris, a dataset specifically designed for predicting evaluation scores under direct supervision from human-annotated judgments. As part of this work, the authors propose a multimodal metric learning framework based on human feedback, which handles both image and text inputs and learns directly from human evaluations based on multimodal inputs.

Building on this foundation, the DENEB metric [Matsuda et al., 2024] introduces a supervised evaluation approach specifically designed to be robust against hallucinations. DENEB processes multiple reference captions simultaneously to effectively capture similarities between an image, a candidate caption, and reference captions. Furthermore, to improve the visual diversity of the Polaris dataset, DENEB introduces Nebula, an extended dataset containing approximately three times the number of images, further enhancing its robustness and scalability for evaluation tasks.

Fine-grained Oriented Metrics. More recent metrics still utilize multimodal models (e.g. CLIP) while integrating additional components to enhance performance, particularly by improving their ability to reward fine-grained details.

Some metrics adopt a reference-free approach, arguing that reliance on reference captions introduces challenges and additional costs, thereby increasing the complexity of the evaluation process. For instance, BRIDGE [Sarto et al., 2024a] is a reference-free metric that employs a dual-encoder architecture combined with a mapping module responsible for generating multimodal pseudo-captions, which integrate textual and dense visual features. Similarly, InfoMetIC [Hu et al., 2023] is designed to provide fine-grained feedback on caption quality. Beyond assessing precision and recall, it identifies specific errors, such as incorrect words and unmentioned image regions. To achieve this, InfoMetIC incorporates a fusion module to model intra- and inter-modality relationships, along with a fine-grained module that enhances the accuracy of error localization in both text and image content.

Available in both reference-free and reference-based versions, HICE-S [Zeng et al., 2024] highlights the limitations of CLIP-based metrics, which primarily assess global imagetext compatibility but often struggle with detecting local textual hallucinations and maintaining sensitivity to small visual objects. To overcome these challenges, HICE-S introduces a hierarchical scoring mechanism that leverages the SAM model [Kirillov et al., 2023] to generate masks for localized visual regions and corresponding textual phrases. These are then processed through a modified version of the CLIP architecture [Sun et al., 2024], enhancing the model's ability to capture fine-grained visual-semantic relationships.

Following a similar hierarchical approach, HiFi-S [Yao et al., 2024] is a fine-grained image description evaluation metric that represents both text and images as parsing graphs. These graphs organize multi-granular instances into a hierarchical structure based on their inclusion relationships, enabling a comprehensive scene analysis across modalities from global to local levels. Additionally, HiFi-S incorporates an LLM to evaluate the fluency of candidate descriptions, further enhancing the evaluation process.

2.3 LLMs-basedMetrics

In recent years, the integration of LLMs into the captioning evaluation pipeline has gained popularity, as demonstrated by HiFi-S. While first attempts compared generated sentences with reference captions using BERT-based embeddings, more recent approaches leverage the advanced reasoning and extensive pre-training capabilities of LLMs to produce more robust evaluation scores. For example, CLAIR [Chan et al., 2023] utilizes an LLM to rate the alignment of a candidate caption with a set of reference captions. Notably, CLAIR evaluates captions without considering image content. In response, FLEUR [Lee et al., 2024] is a reference-free, explainable evaluation metric for image captioning, which directly leverages a score generated by an MLLM and adjusted with a smoothing function to better align with human judgments.

2.4 Hallucination-orientedMetrics

In image captioning, hallucination refers to the inclusion of information, objects, or details in a generated caption that are not present in the corresponding image, compromising the reliability and quality of the description. Accurately identifying captions with potential object hallucinations is therefore essential. Metrics such as CHAIR [Rohrbach et al., 2018] and ALOHa [Petryk et al., 2024] are specifically designed to address this challenge. Specifically, the CHAIR metric calculates what proportion of words generated is actually in the image according to the ground-truth sentences and detected object. However, this metric is restricted to a fixed set of COCO objects and their synonyms. To overcome this limitation, ALOHa introduces an open-vocabulary approach that utilizes LLMs to detect object hallucinations. Specifically, ALOHa extracts groundable objects from the candidate caption using an LLM, measures their semantic similarity to reference objects in captions or detections, and applies Hungarian matching to compute the final hallucination score.

3 Experimental Evaluation

This section presents a comprehensive analysis of captioning evaluation metrics through quantitative experiments. The evaluation spans multiple dimensions, including correlation with human judgments, ranking accuracy, and sensitivity to object hallucinations. Experiments are conducted on diverse datasets, with results summarized in Table 1. Additionally, we assess how well existing metrics evaluate captions generated by modern MLLMs, with results presented in Table 2.

3.1 Correlation with Human Judgment

We analyze the correlation of metrics with human judgment, highlighting their varying behaviours. Experimental results are reported on standard captioning evaluation datasets, including Flickr8k-Expert, Flickr8k-CF, and Composite, along with the more recent Polaris and Nebula datasets.

Datasets. Flickr8k-Expert [Hodosh et al., 2013] contains 17k annotations for 5,664 images, with each pair scored from 1 (no correlation) to 4 (accurate depiction). Flickr8k-CF [Hodosh et al., 2013] provides 145k binary quality judgments for 48k pairs across 1,000 unique images, using the mean proportion of "yes" votes as the alignment score. The Composite dataset [Aditya et al., 2015] includes 12k human ratings for image-caption pairs (with around 4k unique images), assessed on a 1-5 scale. However, these datasets lack modelgenerated captions, leading to a domain gap when applying them to train evaluation metrics. To address this limitation, the Polaris dataset [Wada et al., 2024] introduces 131k human judgments from 550 evaluators (i.e., around ten times larger than previous datasets) covering both human-written and machine-generated captions from ten image captioning models. Expanding on Polaris, the Nebula dataset [Matsuda et al., 2024] includes 32,978 images with human judgments from 805 annotators. In line with prior studies [Hessel et al., 2021; Wada et al., 2024], we use Kendall's correlation coefficient τb for Flickr8k-CF and Kendall's τc for the other datasets.

Experimental Analysis and Key Takeaways. Among rulebased methods, CIDEr demonstrates the highest performance across most datasets, with the exception of Flickr8k-Expert and Composite, where SPICE outperforms it by +1 and +2.6 points, respectively. Within the learnable metrics, RefPACS++ (ViT-L) achieves the best results on Flickr8k-Expert, Flickr8k-CF, and Polaris, while ViLBERT-S yields the highest scores among metrics evaluated on the Nebula dataset. Notably, reference-based metrics generally outperform their reference-free counterparts. However, among the referencefree methods, InfoMetIC and PAC-S variants demonstrate superior performance, emphasizing the importance of refining large pre-trained backbones in the absence of reference captions. When scaling the backbone size, PAC-S variants maintain superior performance, demonstrating their robustness and scalability. In a supervised setting, comparisons are more complex due to differences in backbone sizes. Overall, DENEB (ViT-L) achieves the highest results across multiple datasets. Interestingly, despite leveraging a supervised approach and incorporating an additional learnable module, DENEB is outperformed by RefPAC-S++ (ViT-L), which is trained in an unsupervised manner. This highlights the effectiveness of leveraging CLIP pre-training and enhancing it with regularization through additional generated visualtextual pairs. For fine-grained evaluation, the HiFi metric, employing a hierarchical parsing graph, achieves the best results in a reference-free setting, surpassing HICE-S by +7.9 and BRIDGE (ViT-L) by +8.6 on the Composite dataset. On the Flickr8k-CF dataset, the best performance is instead achieved by RefHICE-S. Among LLM-based metrics, FLEUR performs best, highlighting the importance of leveraging a multimodal approach that incorporates input images, unlike the CLAIR metric which relies solely on reference captions for evaluation. This underscores the critical role of visual information in the captioning task.

Overall, no single metric consistently outperforms all others across datasets. Interestingly, on the Flickr8k-CF dataset, comparable results are achieved using both RefPAC-S++ and the MLLM-based metric FLEUR. This comparison is particularly noteworthy, as the FLEUR metric, based on a LLaVA model with 13B parameters, achieves similar performance to metrics that utilize a smaller CLIP model with around 400M parameters. This highlights that for specific tasks like captioning evaluation, refining a pre-trained embedding space may be more effective than relying on a larger, multi-task embedding. Moreover, there is a significant performance gap between rule-based metrics and newer approaches, highlighting the importance of leveraging input images and the advantages of large-scale pre-training in achieving superior performance.

3.2 Pairwise Ranking

We focus on the pairwise ranking ability of current captioning metrics, measuring performance on the Pascal-50S dataset.

Dataset. Pascal-50S [Vedantam et al., 2015] evaluates captioning metrics using pairwise preference judgments. It consists of 4,000 sentence pairs linked to 1,000 images, each with 50 reference captions. Human judges determine which caption better describes the image, categorizing pairs into human-correct, human with one incorrect caption, human vs. machine-generated, and machine-generated. In this setting, we compute accuracy instead of correlation scores, reporting the averaged results across the four categories.

Experimental Analysis and Key Takeaways. Among rulebased metrics, METEOR achieves the highest accuracy, outperforming CIDEr by +1 point. For learnable unsupervised metrics, the results deviate from trends seen in correlationbased evaluations, where larger backbones and referencebased metrics typically excel. In this case, InfoMetIC, a reference-free metric using the ViT-B backbone, emerges as the best among all metrics. Its ability to identify incorrect semantic words and unmentioned visual regions proves advantageous for this task. Conversely, for other learnable and LLM-based metrics, the results closely align with those observed in human correlation evaluations, with the highest overall accuracy achieved by DENEB (ViT-L).

3.3 Sensitivity to Object Hallucinations

We evaluate robustness to hallucinations on the FOIL dataset.

Dataset. FOIL [Shekhar et al., 2017] consists of imagecaption pairs derived from COCO [Lin et al., 2014], where captions are intentionally altered by introducing a single erroneous word, that makes the modified caption closely resembles the original but includes a specific mistake. In our evaluation, we compute the percentage of times the original caption gets the highest score, according to each metric.

Experimental Analysis and Key Takeaways. In this task, CIDEr outperforms all other rule-based metrics by a substantial margin, with a gain of around +10 points. Among learnable metrics, RefHICE achieves the highest accuracy, followed by DENEB and RefPAC-S, both using the ViT-L backbone. Reference-free versions of HICE and PAC-S experience a significant drop, indicating the need for reference captions to detect hallucinated objects. The results also highlight the advantages of stronger backbones. For instance, PAC-S (ViT-L) outperforms PAC-S (ViT-B) by +2 points, demonstrating the impact of model architecture. Notably, rule-based metrics like CIDEr are significantly surpassed by modern multimodal approaches, with a performance gap of approximately 10 points compared to RefPAC-S (ViT-L) and nearly 12 points against the LLM-based RefFLEUR. This disparity highlights the limitations of rule-based approaches in effectively capturing the nuanced semantics of image captions, particularly in detecting subtle errors such as hallucinations. Metrics like RefPAC-S and RefFLEUR benefit not only from powerful backbones but also from their ability to align text with visual content, allowing them to detect discrepancies much more accurately than traditional metrics. These results empathize the growing need for evaluation metrics that assess not just linguistic fluency but also semantic accuracy and visual relevance through advanced multimodal understanding.

3.4 System-level Correlation

We assess whether well-established captioning metrics effectively evaluate captions generated by modern MLLMs, especially when their style deviates from traditional COCO captions. The potential limitations of these metrics originate from their design: rule-based metrics rely on reference captions that adhere to COCO-style annotations, while learnable metrics are primarily fine-tuned on COCO or similar datasets. As a result, their ability to evaluate captions with diverse linguistic structures remains uncertain. To investigate this, we assess MLLM-generated captions in two formats: short captions, comparable in length to COCO ones, and longer, unconstrained descriptions 1 . This enables us to analyze metric performance across varying caption styles and lengths.

1 For short captions, we use the prompt "Briefly describe the image". For longer ones, we rely on the MLLM default prompt (e.g. "What is the content of the image?").

In Table 2, we evaluate popular captioning models on the COCO test set using a range of metrics to assess their effectiveness and highlight differences in their behavior 2 . Our analysis includes both traditional captioning models and modern MLLMs designed for broader tasks. Additionally, we establish a human-based baseline by randomly selecting one human-annotated caption from the five provided in COCO and comparing it against the remaining references 3 . For evaluation, we employ standard rule-based scores such as BLEU, METEOR and CIDEr, as well as two widely used learnable metrics, CLIP-S and PAC-S++, along with their referencebased counterparts. Additionally, we include two recent evaluation methods: Polos, a supervised metric, and BRIDGE, which focuses on fine-grained visual details. This selection allows for assessing how different evaluation methods capture caption quality across various models.

As it can be seen, for models explicitly trained for the image captioning task, such as M2 Transformer, the generated captions tend to align with COCO in both length and style. Under these conditions, traditional evaluation metrics like CIDEr, effectively recognize the superior quality of BLIP2 captions. Similarly, recently developed metrics leveraging large-scale pre-trained models maintain strong performance.

When evaluating captions generated by MLLMs interesting patterns emerge. If the captions remain similar in length to the ones contained in COCO, rule-based metrics still perform effectively: CIDEr, for instance, assigns a score of 142.2 to captions generated by InstructBLIP. However, as caption length increases, these metrics exhibit a sharp decline in scores. Notably, when evaluating LLaVA-1.5, which produces significantly longer captions than COCO ones, CIDEr score drops dramatically to just 0.3. Qualitative results showing this trend are reported in Fig. 2. This decline stems from rule-based metrics relying on COCO-style captions, which are shorter and simpler than detailed MLLM outputs, resulting in misleading evaluations.

2 Note that for BLIP-2 we employ the captioning-specific model fine-tuned on COCO image-caption pairs.

3 BLEU is not reported for the human-based baseline as its value is sensitive to the number of reference captions.

In contrast, more recent metrics that incorporate largescale components exhibit greater robustness to variations in caption length. A slight decline in performance is still observed: for example, PAC-S++ decreases from 0.710 to 0.697 when caption length increases nearly ninefold (i.e., for IDEFICS-3). This primarily reflects a lower confidence level rather than a failure of the metric. In fact, although these metrics were not explicitly designed for evaluating long, highly detailed captions, they still maintain a high degree of reliability. For learnable reference-based metrics such as RefCLIPS, RefPAC-S++, and Polos, performance degrades more significantly compared to their reference-free counterparts, reflecting a trend similar to that observed in rule-based metrics. However, unlike rule-based solutions, these metrics retain their ability to distinguish high-quality captions, correctly rewarding models such as Llama-3.2 and InstructBLIP.

Overall, these results indicate that recent learnable metrics designed for image captioning remain valuable for evaluating longer and more detailed captions generated by MLLMs, demonstrating reliability despite variations in length and style. However, reference-free metrics are generally preferable, as they more effectively distinguish high-quality captions and exhibit greater robustness to variations in length.

4 Future Directions and Open Problems

Despite progress in image captioning evaluation, several challenges remain, offering opportunities for future research.

Benchmark Evolution. Traditional evaluation datasets mainly feature short captions similar to those in COCO, creating a gap with the longer, more detailed outputs of modern MLLMs. While recent efforts, like Polaris and Nebula, incorporate captions from standard models, they often maintain the concise style of COCO descriptions. A valuable direction would involve creating new benchmarks specifically designed to assess the quality of metrics on longer, richer captions that better reflect the MLLM outputs, addressing challenges like synonyms, paraphrases, and domain-specific terms.

Explainability in Metrics. Most evaluation metrics generate scores without offering insights into the rationale behind their assessments, limiting their usefulness for improving captioning models. A few metrics, such as CLAIR, have advanced in this area by leveraging the interpretative capabilities of LLMs to offer explanations alongside scores. Building upon these advancements, future research should focus on further enhancing explainability, facilitating a deeper understanding of the strengths and weaknesses of evaluation models.

Detecting Hallucinations. Hallucination remains a major challenge in captioning, with models often generating inaccurate details. While benchmarks like FOIL help identify simpler errors, modern MLLMs produce more complex hallucinations, such as distorted relationships or invented elements. To address this, more diverse datasets are needed to assess these complex errors, including test cases designed to challenge evaluation metrics on altered attributes, incorrect relationships, or non-existent entities and actions. Additionally, although some metrics targeting hallucinations have emerged, more advanced solutions are required to capture the diversity and semantic richness of MLLM-generated captions.

Metrics Personalization. Current evaluation metrics rely on reference captions or image-caption alignment, but they fail to accommodate the diversity of user preferences or taskspecific requirements. Future research should focus on the development of personalized evaluation metrics, that can let users prioritize factors such as detail, brevity, stylistic preferences, or domain relevance. By incorporating these custom priorities, evaluation can become more adaptable and meaningful across a variety of applications and user needs.

5 Conclusion

This survey examines image captioning evaluation metrics with the advent of MLLMs, comparing their performance on standard datasets and analyzing their adaptability to evolving caption styles. Our analysis reveals that traditional rulebased metrics struggle to evaluate longer, more detailed captions, while learnable metrics remain effective. Moreover, we highlight key open challenges and promising future directions, emphasizing the need for more robust, context-aware evaluation frameworks that can assess the growing diversity and complexity of captioning models.

Acknowledgments

We acknowledge the CINECA award under the ISCRA initiative, for the availability of high-performance computing resources. This work has been conducted under a research grant co-funded by Leonardo S.p.A. and supported by the PNRR-M4C2 project FAIR and the PRIN 2022-PNRR project MUCES (CUP E53D23016290001), funded by EU - Next-Generation EU. The authors also thank Davide Caffagni and Lorenzo Baraldi for their valuable support.