RAVEN++: Pinpointing Fine-Grained Violations in Advertisement Videos with Active Reinforcement Reasoning

Deyi Ji, Yuekui Yang, Liqun Liu, Peng Shu, Haiyang Wu, Suhua Tang, Xudong Chen, Shaoping Ma, Tianrun Chen, Lanyun Zhu

Abstract

Citation: Deyi Ji, Yuekui Yang, Liqun Liu, Peng Shu, Haiyang Wu, Shaogang Tang, Xudong Chen, Shaoping Ma, Tianrun Chen, Lanyun Zhu. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track. 2025.

RAVEN++ is a specialized multimodal framework designed to moderate video advertisements by identifying not just if a violation occurs, but exactly where and why. By combining hierarchical reinforcement learning with an "active" training strategy that prioritizes difficult samples, it significantly outperforms general-purpose LLMs and previous state-of-the-art models in precision, temporal grounding, and explainability.

Introduction

In the digital economy, video advertisement moderation is a high-stakes task. Platforms must enforce strict policies to ensure legal compliance and maintain a positive user experience. However, video ads are notoriously difficult to moderate because violations are often subtle, brief, and context-dependent.

While previous models like the original RAVEN improved coarse-grained detection (e.g., "Does this ad contain violence?"), they struggled with three critical areas:

  1. Granularity: Distinguishing specific sub-categories (e.g., "Gory Content" vs. general "Discomforting Content").
  2. Localization: Pinpointing exact timestamps (sub-scenes) where violations occur.
  3. Explainability: Aligning "Chain-of-Thought" reasoning with final verdicts without verbosity or contradiction.

Methodology

RAVEN++ evolves the moderation process through a three-stage progressive training pipeline designed to enhance both knowledge and reasoning capabilities.

RAVEN++ Three-Stage Training Pipeline

Stage 1: Knowledge Injection (SFT)

The model is initialized using Supervised Fine-Tuning (SFT). It "learns the manual" by processing ad content summaries and Q&A pairs generated by a larger teacher model (Qwen-72B) regarding specific moderation rules.

Stage 2: Curriculum-Based Passive RL

Using Group Relative Policy Optimization (GRPO), the model is trained on large-scale noisy data. A hierarchical reward system consists of six components, including:

  • Tversky Distance Rewards: A specialized metric to handle imbalanced datasets by flexibly weighting false positives and negatives.
  • Reasoning Consistency: A reward penalizing internal logic (Chain-of-Thought) that deviates from high-quality reasoning distilled from the teacher model.
  • Temporal Grounding: Rewards for accurately identifying the start and end times of violative scenes.

Stage 3: Active Reinforcement Learning

This core innovation dynamically groups samples by difficulty. If the model fails basic detection, the sample is sent to an SFT buffer for "re-teaching." If the model identifies the major category but misses the sub-category or timing, it is sent to an RL buffer with higher weights to refine its precision.

Experimental Setup

The researchers evaluated RAVEN++ on two primary fronts:

  • Industrial Dataset: A proprietary dataset from Tencent consisting of 38,000 training videos (mixed annotations) and 5,000 precisely annotated test videos.
  • MultiHateClip: A public benchmark for multilingual hateful video detection, testing the model's out-of-domain (OOD) generalization.

Results and Analysis

RAVEN++ demonstrated superior performance across all metrics compared to baselines like LLaVA-v1.5, Qwen2.5-VL, and the original RAVEN.

Performance Comparison on Major Violation Categories

Key results include:

  • Precision and Recall: On the industrial dataset, RAVEN++ achieved an average Major Category Precision/Recall of 0.878/0.828, significantly higher than the original RAVEN (0.826/0.788).
  • Temporal Grounding: The model's ability to locate violations (measured by mIoU) improved from 0.555 to 0.647.
  • Sub-Category Accuracy: It showed a nearly 9% improvement in identifying fine-grained violation types.
  • Real-World Impact: In a day-long online A/B test on a live business platform, RAVEN++ delivered a 9.9% improvement in temporal interval localization accuracy over the previous production model.

Discussion

The authors highlight the shift from "passive" to "active" learning. By choosing which samples need SFT versus RL, the model becomes more data-efficient. Furthermore, treating reasoning as a measurable objective via BERT-based semantic similarity helps mitigate "hallucinations" in explanations, ensuring the logic supports the verdict. The hierarchical approach (Major Category -> Sub-Category -> Rule) allows the model to learn the logical structure of policy enforcement.

Conclusion and Future Work

While RAVEN++ sets a new benchmark for video-based moderation, it currently focuses primarily on the visual modality. Real-world advertisements often contain violations in audio tracks or text overlays that may contradict visual content. Future work will focus on extending the framework into a truly unified multi-modal system capable of analyzing video, audio, and OCR-extracted text simultaneously to catch cross-modal inconsistencies.

Summary updated

BibTeX
@misc{ji2025raven,
  author = {Deyi Ji and Yuekui Yang and Liqun Liu and Peng Shu and Haiyang Wu and Suhua Tang and Xudong Chen and Shaoping Ma and Tianrun Chen and Lanyun Zhu},
  title = {{RAVEN++: Pinpointing Fine-Grained Violations in Advertisement Videos with Active Reinforcement Reasoning}},
  year = {2025},
  eprint = {2511.19168},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.19168},
}
Figure