Title: Perception First: A Frontier Native-Video Model with Self-Consistency for Implicit Video Question Answering

Abstract:

This paper outlines our entry into the VRR Challenge @ CVPR 2026, which utilizes the \emph{ImplicitQA} and \emph{VRR-QA} benchmarks~\cite{implicitqa}. These benchmarks involve multiple-choice video question answering tasks where the correct answers are intentionally unobservable in any single frame. Instead, solutions must be deduced by analyzing spatial layout, motion, depth, viewpoint, causality, and social context across discontinuous segments of creative video.

We performed a comprehensive, training-free evaluation of open-source Video-LMMs—including Qwen2.5-VL~\cite{qwen25vl}, Qwen3-VL~\cite{qwen3vl}, InternVL3, Gemma-3, and RL-tuned reasoners such as Video-R1~\cite{videor1} and VideoChat-R1.5~\cite{videochatr15}—combined with various inference-time strategies. These strategies included chain-of-thought, question decomposition, describe-then-reason cascades, audio transcripts, spatial state prompting, self-consistency~\cite{selfconsistency}, multi-model ensembling, and category routing.

Our primary conclusion is that this benchmark is constrained by perception capabilities rather than reasoning limitations. Consequently, augmentations focused on reasoning proved either neutral or detrimental, while the base model’s perceptual strength and lightweight test-time denoicing emerged as the only effective methods for improvement. Detailed error analysis by category identified low-level perception tasks—specifically relative depth, viewpoint, and counting—as the most difficult, whereas causal and social reasoning were nearly mastered. Notably, a prompt designed to explicitly inject monocular depth cues to address the weakest category actually reduced test accuracy by $5.8$ points. This result confirms that the models require enhanced \emph{perception} rather than refined \emph{procedures}.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC