FADE: From Passive Verification to Active Discovery in Counterfactual Video Understanding

2026-08-11

AI Summary

This work addresses a critical limitation in existing counterfactual video understanding benchmarks, where questions and answer choices inadvertently reveal the target event, enabling models to merely verify rather than actively discover anomalies. To overcome this, the authors propose FADE, a novel training framework employing an evidence-first, two-stage paradigm: first, evidence-internalization supervised fine-tuning anchors key visual anomalies; second, fading-anchor reinforcement learning progressively removes textual cues, compelling the model to independently identify and explain counterfactual evidence. FADE establishes the first training mechanism that shifts from passive verification to active discovery and introduces a multi-task alignment evaluation pipeline requiring no additional annotations, uniformly supporting multiple-choice questions, open-ended QA, and description tasks. It achieves state-of-the-art performance across all three tasks on DualityVidQA-test and IPV-Bench, with 90.4% and 67.4% performance retention when transferring from MCQ to open QA and description tasks, respectively—significantly outperforming GPT-5.6.

Abstract

Counterfactual video understanding evaluates whether models grasp physical and commonsense regularities. However, existing multiple-choice question (MCQ) benchmarks inadvertently leak target events through their questions and candidate options. This reduces the core challenge from active discovery to text-guided verification. In this paper, we present FADE, an effective training framework for counterfactual discovery and explanation. Our method is built on an evidence-first, two-stage training paradigm. First, evidence-internalized supervised fine-tuning grounds the model's predictions in decisive visual anomalies. Second, we apply a fading-anchor reinforcement learning strategy that progressively removes textual guidance, compelling the model to independently discover and explain evidence. To rigorously evaluate this capability, we also introduce an effective pipeline that converts existing MCQ datasets into aligned MCQ, open-ended question answering (OQA), and captioning tasks without requiring additional data curation. Our simple approach yields strong results. Using Qwen3-VL-8B as the baseline, FADE achieves state-of-the-art strict paired scores across all three tasks on DualityVidQA-test and IPV-Bench, outperforming GPT-5.6. In specific, when transitioning from constrained MCQs to unconstrained OQA and captioning, our model demonstrates remarkable robustness. Its performance retention is 90.4% and 67.4% on DualityVidQA-test-substantially higher than the 48.1% and 30.7% retained by GPT-5.6. We hope this simple framework can serve as a solid baseline for future research in unconstrained counterfactual video understanding.

Problem

• counterfactual video understanding
• multiple-choice question benchmarks
• active discovery
• text-guided verification
• visual anomalies

Innovation

• counterfactual video understanding
• active discovery
• evidence-first training
• fading-anchor reinforcement learning
• unconstrained evaluation