Concrete Jungle: Towards Concreteness Paved Contrastive Negative Mining for Compositional Understanding
Summary
Proposes Slipform, a training framework that uses lexical concreteness to select harder negatives and a margin-based Cement loss, boosting compositional reasoning in vision-language models.
View Cached Full Text
Cached at: 04/21/26, 11:27 AM
Paper page - Concrete Jungle: Towards Concreteness Paved Contrastive Negative Mining for Compositional Understanding
Source: https://huggingface.co/papers/2604.13313
Abstract
Vision-language models face challenges in compositional reasoning due to insufficient samples for distinguishing subtle semantics, which are addressed through lexical concreteness-based negative sample selection and a novel margin-based loss function.
Vision-Language Modelsdemonstrate remarkable capabilities but often struggle withcompositional reasoning, exhibiting vulnerabilities regarding word order and attribute binding. This limitation arises from a scarcity of informative samples needed to differentiate subtle semantic variations duringcontrastive pretraining. Althoughhard negative miningoffers a promising remedy, existing methods lack explicit mechanisms to dictate which linguistic elements undergo modification. Instead of engineering generative architectures, this study establisheslexical concretenessas a fundamental determinant of negative sample efficacy. Modifying highly concrete terms generates more pronounced structural and visual discrepancies, providing a substantially stronger learning signal. Leveraging this principle, ConcretePlant is proposed to systematically isolate and manipulate perceptually grounded concepts. Analyses of theInfoNCEfurther reveals a severegradient imbalance, where easily distinguishable pairs disproportionately overwhelm the optimization process and restrict the bandwidth available for nuanced learning. To resolve this degradation, theCement lossis formulated utilizing a margin-based approach. By correlating psycholinguistic scores with sample difficulty, this objective dynamically calibrates the penalization applied to individual training pairs. Comprehensive evaluations substantiate these theoretical claims. The integrated framework, designated asSlipform, achieves state-of-the-art accuracy across diverse compositional evaluation benchmarks, general cross-modal retrieval, single and multi label linear probing.
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2604\.13313
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2604.13313 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2604.13313 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2604.13313 in a Space README.md to link it from this page.
Collections including this paper1
Similar Articles
HNC: Leveraging Hard Negative Captions towards Models with Fine-Grained Visual-Linguistic Comprehension Capabilities
The paper introduces Hard Negative Captions (HNC), a dataset and method for training vision-language models to achieve fine-grained comprehension by addressing weak associations in web-collected image-text pairs.
Adversarial Concept Search: Predicting Compositional Errors From Feature Geometry
This paper proposes Adversarial Concept Search, a method that uses the representational geometry of large language models to predict compositional failures without evaluating specific inputs. The approach identifies high-risk scenarios by measuring interference between salient features.
CopT: Contrastive On-Policy Thinking with Continuous Spaces for General and Agentic Reasoning
CopT introduces a contrastive on-policy thinking framework for LLMs that generates draft answers first, then uses contrastive verification and dynamic thinking to improve accuracy while reducing token consumption, achieving up to 23% higher accuracy and 57% lower token usage on math, coding, and agentic reasoning tasks.
COMPASS: Grounding Composition-Intent Guidance in Unified Multimodal Models
This paper presents COMPASS, the first unified multimodal framework that grounds composition-intent control for both composition perception and composition-guided generation, introducing a shared expert token and the Comp-11 dataset.
HELP WITH RESEARCH: Observation - Semantically Dense Context Produces Strong Late-Layer Divergence Without Jailbreak Prompts [D]
An empirical study demonstrating that long, semantically dense, benign text can shift a model's latent space and bypass alignment, causing it to generate otherwise blocked critiques. The author, a non-expert, requests an audit of their metrics to distinguish genuine semantic hijacking from artifacts.