VITED: Video Temporal Evidence Distillation

2025-03-17

AI Summary

Existing video question answering (VQA) models uniformly sample a fixed number of frames, hindering the capture of temporally sparse yet critical evidence and lacking the capability to localize multiple disjoint evidence segments across the full video context for multi-step reasoning. Method: We propose a Chain-based Temporal Evidence Reasoning framework that pioneers automatic construction and end-to-end generation of video evidence chains, breaking away from fixed-frame sampling. Our approach integrates optimization-driven evidence interval discovery, evidence-chain supervised training, and joint modeling of temporal localization and multi-hop reasoning. Contribution/Results: The framework enables fine-grained, cross-temporal visual evidence localization and compositional reasoning. It significantly outperforms state-of-the-art (SOTA) evidence-agnostic VQA methods on multiple long-video QA benchmarks, achieving substantial improvements in answer accuracy—particularly for complex, multi-step questions.

Technology Category

Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal ReasoningComputer Vision: Video Understanding & Activity AnalysisReasoning under Uncertainty: Other Foundations of Reasoning under Uncertainty

Application Category

Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAGGraph Algorithms and Modeling for the Web: Algorithms and analysis for heterogeneous, signed, attributed, multi-relational, temporal, higher-order, and annotated Web-related graphsSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

We investigate complex video question answering via chain-of-evidence reasoning -- identifying sequences of temporal spans from multiple relevant parts of the video, together with visual evidence within them. Existing models struggle with multi-step reasoning as they uniformly sample a fixed number of frames, which can miss critical evidence distributed nonuniformly throughout the video. Moreover, they lack the ability to temporally localize such evidence in the broader context of the full video, which is required for answering complex questions. We propose a framework to enhance existing VideoQA datasets with evidence reasoning chains, automatically constructed by searching for optimal intervals of interest in the video with supporting evidence, that maximizes the likelihood of answering a given question. We train our model (VITED) to generate these evidence chains directly, enabling it to both localize evidence windows as well as perform multi-step reasoning across them in long-form video content. We show the value of our evidence-distilled models on a suite of long video QA benchmarks where we outperform state-of-the-art approaches that lack evidence reasoning capabilities.

Problem

• Enhance video question answering via evidence reasoning chains
• Localize critical evidence in long-form video content
• Improve multi-step reasoning in complex video QA tasks

Innovation

• Chain-of-evidence reasoning for video QA
• Automated optimal interval search for evidence
• Direct generation of evidence chains by VITED