arXiv Computer Vision By Linyuan Gao, Yuan Wu, Yi Chang

CCRV-Bench: Constraint-Based Evaluation of Causal Reasoning in Vision-Language Models

Read the original on arXiv Computer Vision →

CCRV-Bench is a constraint‑driven benchmark designed to evaluate visual causal reasoning in vision‑language models on single‑image physical scenarios. It assesses four causal task dimensions—causal relation discovery, state prediction, causal diagnosis, and intervention—while applying constraints such as entity symbolization, spatial grounding, factual adversarial constraints, and minimalist output constraints to reduce shortcut learning. Experiments on 15 multimodal models reveal that constraint sensitivity varies by task and model, with intervention and spatial grounding having the largest impact and factual adversarial constraints improving causal diagnosis across models.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Computer Vision.

arXiv AI
6d ago

Seeing Isn't Enough: Unlocking Visual Reasoning by Bridging the Reasoning-Access Gap with TIGER

The paper introduces TIGER, a training framework that addresses the modality reasoning‑access gap in multimodal large language models (MLLMs). By generating synthetic counterfactuals where models succeed on text but fail on rendered images, TIGER fine‑tunes MLLMs using Group Relative Policy Optimization, SFT, and DPO, improving performance on real‑world visual benchmarks such as MathVerse, EMMA, MMMU, ChartQAPro, and RoboSpatial‑Home. The approach demonstrates that visual reasoning can be enhanced without multimodal annotations, and it complements existing RLVR methods.

By Gregory Kang Ruey Lau, Minh Huynh Nguyen, Bryan Kian Hsiang Low
arXiv AI
6d ago

CASCADE: Constraint-Aware SCAffolded DEcomposition for Logical Anomaly Detection

The paper introduces CASCADE, a training‑free framework that enhances logical anomaly detection by using constraint‑specific visual scaffolds instead of text‑only decompositions. CASCADE++ further refines these scaffolds at the object level, achieving state‑of‑the‑art performance on MVTec‑LOCO and VID‑AD with only three reference images and no training. The approach demonstrates that adapting visual input to each constraint significantly improves verification accuracy across multiple vision‑language model backbones.

By Jinchan Kim, Jinhee Im, Hyunsouk Cho