Seed42Lab/RefOI

💬 VLM-REG: Vision-Language Models Are Not Pragmatically Competent in Referring Expression Generation

📃 Paper ｜🏠 Homepage

Overview

Referring Expression Generation (REG)—the task of producing a concise and unambiguous description that allows a listener to identify a target object—lies at the heart of pragmatic communication in vision-language systems. However, existing benchmarks suffer from two major limitations:

1. Data leakage in RefCOCO/RefCOCO+, which raises concerns about evaluation contamination, especially for VLMs trained on MSCOCO.
2. Lack of spoken data, despite the fact that real-world referring is often real-time and spontaneous, unlike written language, which benefits from planning and revision.

To address these gaps, we introduce RefOI, a curated dataset built from the OpenImages V7 Instance Segmentation validation set.

Key features:

• 1,485 real-world object instances, equally distributed across COCO (744) and non-COCO (741) classes.
• Includes single presence and co-occurrence images for each class.
• Each instance annotated with 3 written and 2 spoken human referring expressions.

Using RefOI, we evaluate several state-of-the-art VLMs and uncover three tiers of pragmatic failure:

• Ambiguity: Generated expressions often fail to uniquely identify the referent.
• Redundancy: Models include excessive or irrelevant details, violating principles of informativeness and efficiency.
• Misalignment: Model preferences diverge from human pragmatics, favoring visual complexity over minimal spatial cues.

For token-level annotation of referring expressions, see the companion dataset RefOI-TLHF, which provides minimal span supervision for both human- and model-generated descriptions.

Dataset Schema and Split

Data Fields

Each entry in the dataset contains the following fields:

• image: The original image file.
• mask: A binary segmentation mask isolating the target object.
• distractors: A list of other mask(s) of the class in the image, will be empty for a single_precence entry.
• boxed_image: The original image overlaid with a red bounding box highlighting the target object.
• box_xmin, box_xmax, box_ymin, box_ymax: The normalized bounding‑box coordinates.
• is_coco: A binary flag (1 for COCO-class, 0 for non‑COCO).
• label_name: The object’s class label (e.g., “muffin,” “giraffe”).
• co_occurrence: The number of same‑class instances in the image (1 = no distractors; >1 = multiple).
• written_descriptions: Three human‑typed referring expressions.
• spoken_descriptions: Two human‑spoken expressions (transcribed and optionally corrected by annotators).

Dataset Split

• single_presence (co_occurrence = 1):
  Only one object of the target class appears (no same‑class distractors in the image).

• co_occurrence (co_occurrence > 1):
  Multiple objects of the same class appear in the image, introducing potential referential ambiguity.

Usage

from datasets import load_dataset

# only one object of the class
ds_single = load_dataset("Seed42Lab/RefOI", split="single_presence")
# multiple objects of the class
ds_multi = load_dataset("Seed42Lab/RefOI", split="co_occurrence")

print(ds_single[0])
print(ds_multi[0])

Experiments

We compare multiple models across standard metrics, listener-based accuracy, and human judgment. Humans outperform all models by large margins (e.g., >90% vs. ~50%). Automatic metrics such as BLEU and CIDEr show poor correlation with human judgment, frequently ranking verbose models higher. Even listener-based scores (REC) fail to consistently match human preferences, indicating that existing metrics do not capture pragmatic competence effectively.

Model	Instr.	BLEU-1	BLEU-4	ROUGE-1	ROUGE-L	METEOR	CIDEr	SPICE	BERT	CLIP	REC	Human	Irrel%
LLaVA-7B	Dft.	13.27	1.60	18.09	16.30	19.29	2.10	10.50	85.51	79.02	32.41	39.46	87.30
	Brf.	28.74	6.05	36.46	35.50	19.15	10.80	24.59	89.02	70.72	25.51	30.57	41.95
LLaVA-13B	Dft.	8.17	1.07	11.98	10.94	16.89	0.77	7.92	84.61	79.85	30.13	46.40	91.85
	Brf.	28.96	5.81	36.44	35.64	20.13	8.14	21.63	88.42	72.99	28.92	32.53	49.65
LLaVA-34B	Dft.	6.29	0.78	9.82	9.11	16.15	0.07	7.61	84.39	79.86	33.42	46.53	92.90
	Brf.	28.55	6.38	32.99	31.67	20.48	9.60	16.50	88.50	74.95	35.24	36.77	56.11
XComposer	Dft.	5.25	0.65	8.38	7.81	14.58	3.10	6.37	84.11	79.86	38.06	52.19	92.81
	Brf.	13.59	2.17	17.77	16.69	19.95	5.52	10.63	85.52	79.66	38.47	51.65	80.36
MiniCPM-V	Dft.	6.38	0.67	9.86	8.78	15.28	0.05	6.30	84.29	80.38	37.93	45.12	92.97
	Brf.	16.03	3.15	19.56	18.19	18.77	6.36	11.16	86.29	78.55	35.04	45.79	72.87
GLaMM	Dft.	15.01	3.32	16.69	16.29	11.49	9.08	3.90	86.42	58.26	5.78	3.84	74.68
	Brf.	18.46	4.45	20.92	20.46	14.18	10.48	4.44	86.65	58.60	5.72	4.85	70.52
CogVLM	Dft.	31.13	8.70	33.89	32.32	23.50	41.62	24.09	89.78	66.54	33.29	26.67	26.39
	Brf.	31.39	8.69	34.70	32.94	24.87	41.41	24.74	90.00	69.15	38.80	33.53	29.88
GPT-4o	Dft.	7.47	0.85	11.61	10.43	17.39	0.03	7.21	84.57	80.81	41.29	59.80	89.81
	Brf.	25.30	5.78	28.76	27.36	19.02	8.17	15.31	88.11	76.58	40.08	51.72	52.75
Human	Spk.	66.18	22.58	70.15	66.45	48.28	112.04	42.35	93.89	71.60	64.56	92.20	9.15
	Wrt.	-	-	-	-	-	-	-	-	70.43	63.69	89.29	7.29

Model performance under different Instr. (Instruction) settings: Dft. (Default) prompt and Brf. (Brief) prompt. All model predictions are evaluated against Human Wrt. (Written) results as the reference texts. We also compute Human Spk. (Spoken) data in comparison with human-written data. Irrel% refers to the percentage of irrelevant words in the referring expression of the examples evaluated as successful.

Recommended Use of Our Dataset

The RefOI dataset is designed for fine-grained REG/REC analysis. It distinguishes between COCO and non-COCO classes, and between scenes with single presence vs. co-occurrence of the same class. We encourage users to leverage these distinctions for deeper insights and invite community contributions to expand non-COCO annotations.

Citation

If you find our dataset helpful, please cite our work:

@misc{ma2025visionlanguagemodelspragmaticallycompetent,
      title={Vision-Language Models Are Not Pragmatically Competent in Referring Expression Generation}, 
      author={Ziqiao Ma and Jing Ding and Xuejun Zhang and Dezhi Luo and Jiahe Ding and Sihan Xu and Yuchen Huang and Run Peng and Joyce Chai},
      year={2025},
      eprint={2504.16060},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2504.16060}, 
}