When and why vision-language models behave like bags-of-words, and what to do about it?
read the original abstract
Despite the success of large vision and language models (VLMs) in many downstream applications, it is unclear how well they encode compositional information. Here, we create the Attribution, Relation, and Order (ARO) benchmark to systematically evaluate the ability of VLMs to understand different types of relationships, attributes, and order. ARO consists of Visual Genome Attribution, to test the understanding of objects' properties; Visual Genome Relation, to test for relational understanding; and COCO & Flickr30k-Order, to test for order sensitivity. ARO is orders of magnitude larger than previous benchmarks of compositionality, with more than 50,000 test cases. We show where state-of-the-art VLMs have poor relational understanding, can blunder when linking objects to their attributes, and demonstrate a severe lack of order sensitivity. VLMs are predominantly trained and evaluated on large datasets with rich compositional structure in the images and captions. Yet, training on these datasets has not been enough to address the lack of compositional understanding, and evaluating on these datasets has failed to surface this deficiency. To understand why these limitations emerge and are not represented in the standard tests, we zoom into the evaluation and training procedures. We demonstrate that it is possible to perform well on retrieval over existing datasets without using the composition and order information. Given that contrastive pretraining optimizes for retrieval on datasets with similar shortcuts, we hypothesize that this can explain why the models do not need to learn to represent compositional information. This finding suggests a natural solution: composition-aware hard negative mining. We show that a simple-to-implement modification of contrastive learning significantly improves the performance on tasks requiring understanding of order and compositionality.
This paper has not been read by Pith yet.
Forward citations
Cited by 26 Pith papers
-
Can VLMs Truly Forget? Benchmarking Training-Free Visual Concept Unlearning
VLM-UnBench demonstrates that prompt-based training-free unlearning in VLMs leaves forget accuracy near the no-instruction baseline except under oracle conditions that reveal the target concept.
-
Open-Vocabulary and Referring Segmentation for 3D Gaussians Using 2D Detectors
GaussDet enables open-vocabulary and referring segmentation in 3D Gaussians by learning instance features and aggregating votes from 2D detectors, improving referential grounding by 16.7% mIoU in zero-shot setting.
-
Evaluating and Enhancing Negation Comprehension in Remote Sensing MLLMs
Remote sensing MLLMs perform poorly on negation tasks with hallucinations and accuracy drops, but the NeFo test-time learning method substantially improves negation understanding and generalizes to unseen tasks using ...
-
Advancing Creative Physical Intelligence in Large Multimodal Models
Introduces MM-CreativityBench for affordance-grounded creative tool use and shows that DPO-based alignment with an affordance knowledge base improves entity and part selection while cutting hallucination errors in LMMs.
-
SMA: Submodular Modality Aligner For Data Efficient Multimodal Learning
SMA uses a submodular mutual information objective on data sets to deliver competitive zero-shot classification and retrieval performance on CLIP benchmarks with only tens of thousands of samples, orders of magnitude ...
-
DCR: Counterfactual Attractor Guidance for Rare Compositional Generation
DCR uses a counterfactual attractor and projection-based repulsion to suppress default completion bias in diffusion models, improving fidelity for rare compositional prompts while preserving quality.
-
Adapting MLLMs for Nuanced Video Retrieval
Text-only contrastive fine-tuning of an MLLM with hard negatives produces embeddings that handle temporal, negation, and multimodal nuances in video retrieval and achieves SOTA performance.
-
WISE: A World Knowledge-Informed Semantic Evaluation for Text-to-Image Generation
Text-to-image models show significant limitations in integrating world knowledge, as measured by the new WISE benchmark and WiScore metric across 20 models.
-
Compositionality Emerges in a Narrow Depth-Connectivity Regime: Architecture Constraints and Solution Manifolds
Compositionality emerges in neural networks only in a narrow depth-connectivity regime, with gradient descent converging to fractured solutions outside it.
-
A Systematic Study of Behavioral Cloning for Scientific Data Annotation
Introduces 9 synthetic annotation tasks and benchmarks for behavioral cloning, finding hierarchical skill learning, scaling benefits, effective multi-task pretraining, and shared internal representations of task phase...
-
See What I Mean: Aligning Vision and Language Representations for Video Fine-grained Object Understanding
SWIM aligns cross-attention maps from object nouns to ground-truth masks during training on the new NL-Refer dataset to enable text-only fine-grained video object understanding in MLLMs.
-
Revealing the Gap in Human and VLM Scene Perception through Counterfactual Semantic Saliency
VLMs exhibit size, center, and saliency biases in scene understanding, relying less on people than humans do, with size bias as a key driver of divergence.
-
When Vision Speaks for Sound
Video MLLMs show an audio-visual Clever Hans effect relying on visual-acoustic correlations rather than audio verification; Thud interventions diagnose it and a 10K-sample preference alignment improves intervention pe...
-
Letting the neural code speak: Automated characterization of monkey visual neurons through human language
Natural language descriptions generated via a closed-loop pipeline with digital twins capture the selectivity of most neurons in macaque V1 and V4, with synthesized images driving 96% of V4 neurons into the top or bot...
-
Letting the neural code speak: Automated characterization of monkey visual neurons through human language
Natural-language descriptions generated and verified through generative models and digital twins capture the selectivity of most neurons in macaque V1 and V4.
-
Birds of a Feather Flock Together: Background-Invariant Representations via Linear Structure in VLMs
Exploiting linear structure in VLM embeddings, a synthetic-data pre-training method yields background-invariant representations that exceed 90% worst-group accuracy on Waterbirds even under 100% spurious correlation w...
-
All in One: A Unified Synthetic Data Pipeline for Multimodal Video Understanding
A unified synthetic data generation pipeline produces unlimited annotated multimodal video data across multiple tasks, enabling models trained mostly on synthetic data to generalize effectively to real-world video und...
-
To See or To Please: Uncovering Visual Sycophancy and Split Beliefs in VLMs
69.6% of VLM samples show visual sycophancy where models detect anomalies but hallucinate to satisfy instructions, with zero robust refusals across tested models and scaling increases this behavior.
-
SPHINX: A Synthetic Environment for Visual Perception and Reasoning
SPHINX generates synthetic visual puzzles for benchmarking LVLMs, where GPT-5 scores 51.1% and RLVR training improves both in-domain and external visual reasoning performance.
-
GRAPE: Let GRPO Supervise Query Rewriting by Ranking for Retrieval
GRAPE applies GRPO to an LLM query rewriter with a corpus-relative ranking reward to improve frozen CLIP retrieval by an average 4.9% Recall@10 on shifted benchmarks without retraining or re-embedding.
-
A Composite Activation Function for Learning Stable Binary Representations
HTAF is a sigmoid-tanh composite that approximates the Heaviside function to allow stable gradient training of binary activation networks, yielding ICBMs with stable discretization and competitive performance on image tasks.
-
AFMRL: Attribute-Enhanced Fine-Grained Multi-Modal Representation Learning in E-commerce
AFMRL uses MLLM-generated attributes in attribute-guided contrastive learning and retrieval-aware reinforcement to achieve SOTA fine-grained multimodal retrieval on e-commerce datasets.
-
Concrete Jungle: Towards Concreteness Paved Contrastive Negative Mining for Compositional Understanding
Using lexical concreteness to guide contrastive negative mining and a new margin-based Cement loss, the Slipform framework reaches state-of-the-art on compositional benchmarks for vision-language models.
-
NSFL: A Post-Training Neuro-Symbolic Fuzzy Logic Framework for Boolean Operators in Neural Embeddings
NSFL adapts t-norms and t-conorms to embedding spaces with NS-Delta and SQO to enable logical operations, reporting up to 81% mAP gains in retrieval tasks.
-
Long Story Short: Disentangling Compositionality and Long-Caption Understanding in Contrastive VLMs
Empirical study shows bidirectional but sensitive relationship between compositionality and long-caption understanding in VLMs, promoted by high-quality grounded data and affected by architectural choices like frozen ...
-
ReasonCLIP-58M: Visually Grounded Commonsense Reasoning Supervision for CLIP
ReasonCLIP-58M applies continual pretraining with visually grounded reasoning captions on 58M examples to improve CLIP-style models on commonsense and compositional reasoning tasks.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.