Research map / Vision-language and multimodal / Multimodal LLMs

BLM-Guard: Explainable Multimodal Ad Moderation with Chain-of-Thought and Policy-Aligned Rewards

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes BLM-Guard, an explainable multimodal ad-moderation VLM trained with rule-driven interleaved CoT cold-start SFT followed by GRPO with a critic-based self-consistency reward, and introduces a risk-taxonomy ad benchmark.

Problem

Fine-grained, policy-driven moderation of multimodal short-video ads (deceptive visuals, speech, subtitles), which general community safety filters and VLM guardrails handle poorly, with explainable reasoning required.

Approach

Rule-anchored interleaved CoT data synthesized from keyframes with a frozen InternVL-3-78B teacher, followed by GRPO with a composite reward (rule correctness, format, and a guide-model critique score, SCA-R) on Qwen2.5-VL-7B.

Key results

• On the BLM-Guard Benchmark, BLM-Guard reaches 0.962 weighted accuracy and 0.914 strict accuracy, versus 0.712 / 0.701 for the Qwen2.5-VL-7B backbone.
• Binary risky-class F1 of 0.969 for BLM-Guard versus 0.680 for Qwen2.5-VL-7B and 0.801 for Qwen2.5-VL-32B.
• Ablation: adding SCA-R on top of Rule-SFT and Rule-RL raises strict accuracy from 0.801 to 0.914 and consistency from 0.781 to 0.845.

Surprising findings

• Ablation shows Full-SFT with 13k samples (0.7925 s-Acc) is comparable to Rule-SFT with 5k samples (0.783), suggesting data scaling gives limited gains in that range.
• Baselines LlavaGuard-7B and LLaVA-Next-Video-7B score near chance on the internal benchmark (s-Acc 0.235 and 0.312), while the Qwen2.5-VL-7B backbone reaches 0.701.

Limitations

• The internal benchmark's size, label distribution, and inter-annotator agreement are not reported in the provided text.
• The consistency score relies on GPT-4o as a judge and reward critic, with no human validation described.
• Several claimed components (multi-task architecture for intra-/cross-modal violations, dynamic sampling details) are described only loosely, and the GRPO description has inconsistencies (e.g., token-level vs. sequence-level, the K_i definition).

Experimental setup

Datasets: BLM-Guard Benchmark LSPD XD-Violence UCF-Crime FakeSV FVC

Benchmarks: BLM-Guard Benchmark LSPD XD-Violence UCF-Crime FakeSV FVC

Metrics: weighted accuracy strict accuracy precision recall F1 consistency score

Baselines: LlavaGuard-7B QwenGuard-7B LLaVA-Next-Video-7B InternVL3-8B Qwen2.5-VL-7B-Instruct Kimi-VL-A3B-Thinking Qwen2.5-VL-32B InternVL3-14B

Builds on: GRPO Chain-of-Thought CLIP Qwen2.5-VL

Closest papers by meaning

• Reinforcing Multimodal Reasoning Against Visual Degradation (NeurIPS 2026 · similarity 0.83)
• [Re] Boosting the Visual Interpretability of CLIP via Adversarial Fine-Tuning (NeurIPS 2026 · similarity 0.83)
• Boosting Reasoning in Large Multimodal Models via Activation Replay (CVPR 2026 · similarity 0.83)
• Revisiting the Data Sampling in Multimodal Post-training from a Difficulty-Distinguish View (AAAI 2026 · similarity 0.83)
• PA-FAS: Towards Interpretable and Generalizable Multimodal Face Anti-Spoofing via Path-Augmented Reinforcement Learning (AAAI 2026 · similarity 0.83)
• Multimodal Policy Internalization for Conversational Agents (ICLR 2026 · similarity 0.83)