CoRE is a weakly supervised framework that learns fine-grained temporal and entity support for perceived risk in driving videos using only coarse video-level judgments. It first trains a video-level predictor, freezes it, and then uses structured interventions over candidate temporal regions or entity tracks to generate graded prediction-effect targets. These targets train a student model that can predict temporal and entity support directly from the original video, enabling fine-grained evidence localization without requiring detailed annotations.
By Kaiser Hamid, Can Cui, Nade Liang
The paper introduces REVEAL, a diagnostic benchmark that stresses Video‑Language Models (VidLMs) on five controlled probes—camera‑motion sensitivity, cross‑frame integration, video sycophancy, language‑only shortcuts, and temporal expectation bias—to assess how well these models encode and use visual evidence. Experiments on 12 VidLMs reveal systematic failures: some visual signals are never reliably encoded, while others are overridden by model priors, leading to performance below chance on several probes that humans solve with high accuracy. Mechanistic probes further pinpoint where and why visual evidence is lost, demonstrating that under assertive prompts a model’s output becomes nearly invariant to real versus random video input, rendering visual evidence causally inert.
By Sethuraman T V, Savya Khosla, Aditi Tiwari, Vidya Ganesh, Rakshana Jayaprakash, Aditya Jain, Vignesh Srinivasakumar, Onkar Kishor Susladkar, Srinidhi Sunkara, Aditya Shanmugham, Rakesh Vaideeswaran, Abbaas Alif Mohamed Nishar, Simon Jenni, Rohan Maheshwari, Derek Hoiem
arXiv:2602.10639v2 Announce Type: replace
Abstract: Video Large Language Models (VideoLLMs) have achieved strong performance on video understanding tasks, yet existing benchmarks evaluate only what m...
By Yuxin Cao, Wei Song, Shangzhi Xu, Jingling Xue, Jin Song Dong
The paper introduces FailureSpot, a label‑efficient method for detecting failures at the timestamp level in vision‑language‑action (VLA) policies. It first generates weak supervision from unlabeled VLA action chunks by identifying abnormal patterns, then employs active learning to annotate only the most uncertain trajectories. Experiments on multiple VLA policies demonstrate improved performance for both timestamp‑level and trajectory‑level failure detection.
By Jie Ma, Zongxi Liu, Yi Zhu
PERSIST redefines shot boundary detection as a task of semantic discrimination, requiring a persistent update of a video’s latent temporal state rather than a transient visual change. It employs a FiLM‑conditioned sinusoidal representation network and a structured discriminator that fuses local change, transient impulse, and return‑to‑trend cues into a single interpretable per‑frame signal. The method achieves comparable recall to leading detectors while significantly reducing false positives from flash, text overlay, and archival artifacts, and it is trained solely on real transitions from ClipShots.
By Tingyu Lin, Christian Stippel, Armin Dadras, Jakob Zenzmaier, Florian Kleber, Wolfgang Aigner, Robert Sablatnig
The paper introduces a new protocol, Mistake Detection Video Question Answering (MD‑VQA), to evaluate whether models can determine if a step in a video follows its description, covering both seen and unseen actions. It proposes a post‑training approach for video‑language models that uses a reward function to highlight discrepancies between instructions and video content. Experiments show this method surpasses zero‑shot, fine‑tuned, and other post‑training baselines, especially on unseen procedures, improving performance by up to 11.6% on EP‑VQA.
By Federico Spurio, Olga Zatsarynna, Lars Doorenbos, Emad Bahrami, Gianpiero Francesca, Juergen Gall