Stories about Vision-Language Models

The release of MemeCULT-1K reveals a key fact: the bottleneck for multimodal models in understanding humor lies not in vision or language, but in culture. A minimal context consistently improves performance, indicating current models possess almost no cultural commonsense. This benchmark pushes 'humor understanding' from general capability evaluation to the more precise dimension of 'cultural pragmatic reasoning', providing a quantitative yardstick for future culture-adapted training.

Traditional speculative decoding in VLMs faces a self-defeating cycle: small drafters cannot afford per-step image processing, leading to vision compression. GLANCE reuses the target model's fused vision-language state and adopts one-pass block generation, proving vision need not be computational overhead, architecturally breaking both vision cost and sequential depth constraints.

Vision-language models exhibit modality asymmetry in context-memory conflicts, favoring text over image in-context information. This reveals late representational alignment hinders cognitive control, and simple reasoning enhancements are insufficient; cross-modal fusion design is needed.

The value of this paper lies not in concept blending itself, but in turning the design field's long-standing intuitive principle of positive and negative space into computable constraints. It implies generative visual tools are moving from generating content to understanding compositional semantics, pushing the boundary between design and AI toward professional aesthetic cognition.

The bias in vision-language measurement of urban change stems mainly from re-photography and preprocessing, not from real street changes. Using over four thousand street-view pairs, the study shows that re-shooting the same street produces perception fluctuations equivalent to two-thirds of the difference between different streets, implying longitudinal comparisons based on single images may conflate noise with actual change. Calibration must be built into urban perception research.

This study shows that vision-language models can handle scale-free physical reasoning but systematically under-respond when required to use given scales to convert to real-world units. The root cause may be imbalanced object scale distributions in training data causing reliance on visual priors rather than reference scales, implying that improving scale utilization may matter more than scaling parameters.

This work breaks the 'model-judges-model' loop, offering the first model-free attribution scheme for VLM evaluation. What matters is not score improvement but the finding that perception deficits cannot be compensated by text prompts alone, shifting focus back to visual encoder fidelity.

Landslide understanding is modeled as a diffusion process that progressively infers presence, extent, and boundaries from noisy latent representations. This unifies detection, segmentation, and trigger interpretation in one probabilistic framework, suggesting vision-language guidance is becoming a viable path from task-specific models to open-world generalist models in remote sensing.

This study reveals through causal perturbation analysis that visual attention faithfulness in vision-language models is not a monolithic attribute but exhibits three heterogeneous modes. This implies that generic attention-based explanation methods may fail, necessitating mode-specific interpretation and debugging strategies, and signals a deepening of interpretability research from NLP into multimodal visual reasoning.