Saved in:
Bibliographic Details
Main Authors: Yang, Yiran, Liu, Zhaowei, Yuan, Yuan, Song, Yukun, Ma, Xiong, Song, Yinghao, Zeng, Xiangji, Sun, Lu, Wang, Yulu, Zhou, Hai, Cui, Shuai, Gong, Zhaohan, Zhang, Jiefei
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2602.18193
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910029841432576
author Yang, Yiran
Liu, Zhaowei
Yuan, Yuan
Song, Yukun
Ma, Xiong
Song, Yinghao
Zeng, Xiangji
Sun, Lu
Wang, Yulu
Zhou, Hai
Cui, Shuai
Gong, Zhaohan
Zhang, Jiefei
author_facet Yang, Yiran
Liu, Zhaowei
Yuan, Yuan
Song, Yukun
Ma, Xiong
Song, Yinghao
Zeng, Xiangji
Sun, Lu
Wang, Yulu
Zhou, Hai
Cui, Shuai
Gong, Zhaohan
Zhang, Jiefei
contents Short-video platforms now host vast multimodal ads whose deceptive visuals, speech and subtitles demand finer-grained, policy-driven moderation than community safety filters. We present BLM-Guard, a content-audit framework for commercial ads that fuses Chain-of-Thought reasoning with rule-based policy principles and a critic-guided reward. A rule-driven ICoT data-synthesis pipeline jump-starts training by generating structured scene descriptions, reasoning chains and labels, cutting annotation costs. Reinforcement learning then refines the model using a composite reward balancing causal coherence with policy adherence. A multitask architecture models intra-modal manipulations (e.g., exaggerated imagery) and cross-modal mismatches (e.g., subtitle-speech drift), boosting robustness. Experiments on real short-video ads show BLM-Guard surpasses strong baselines in accuracy, consistency and generalization.
format Preprint
id arxiv_https___arxiv_org_abs_2602_18193
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle BLM-Guard: Explainable Multimodal Ad Moderation with Chain-of-Thought and Policy-Aligned Rewards
Yang, Yiran
Liu, Zhaowei
Yuan, Yuan
Song, Yukun
Ma, Xiong
Song, Yinghao
Zeng, Xiangji
Sun, Lu
Wang, Yulu
Zhou, Hai
Cui, Shuai
Gong, Zhaohan
Zhang, Jiefei
Computer Vision and Pattern Recognition
Short-video platforms now host vast multimodal ads whose deceptive visuals, speech and subtitles demand finer-grained, policy-driven moderation than community safety filters. We present BLM-Guard, a content-audit framework for commercial ads that fuses Chain-of-Thought reasoning with rule-based policy principles and a critic-guided reward. A rule-driven ICoT data-synthesis pipeline jump-starts training by generating structured scene descriptions, reasoning chains and labels, cutting annotation costs. Reinforcement learning then refines the model using a composite reward balancing causal coherence with policy adherence. A multitask architecture models intra-modal manipulations (e.g., exaggerated imagery) and cross-modal mismatches (e.g., subtitle-speech drift), boosting robustness. Experiments on real short-video ads show BLM-Guard surpasses strong baselines in accuracy, consistency and generalization.
title BLM-Guard: Explainable Multimodal Ad Moderation with Chain-of-Thought and Policy-Aligned Rewards
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2602.18193