Research map / Video / Video-language understanding

Controllable Multi-label Video Safety Detection via Adaptive Tversky Policy Optimization

NeurIPS 2026 · Video › Video-language understanding · approach: Event and anomaly reasoning in video · 0 citations (OpenAlex)

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 ATPO, a GRPO-based RL framework with an Adaptive Tversky Reward that feedback-controls false-positive/false-negative weights to improve multi-label video safety detection and steer precision-recall trade-offs.

Problem

Video safety detection is usually reduced to binary classification with static training objectives and fixed operating points, which ignores the multi-label nature of unsafe videos and the need for controllable precision-recall trade-offs.

Approach

Replaces fixed Tversky/Jaccard coefficients in GRPO rewards with coefficients adapted by a log-ratio feedback controller on EMA-tracked FN/FP counts, targeting a user-specified FN-to-FP ratio globally (G-ATR) or per category (C-ATR).

Key results

• ATPO-G-7B raises SafeWatch-Bench-Real Jaccard from 40.66 (Qwen2.5-VL-7B zero-shot) to 75.44, and reaches 88.17 Jaccard on XD-Violence.
• G-ATR precision-recall ratio rises monotonically from 0.77 to 1.16 as target FN/FP ratio r* goes from 0.2 to 5, whereas static STR gives non-monotonic control.
• Against matched GRPO with STR, ATPO-G improves Real Jaccard by 5.35 points (70.09 to 75.44).

Surprising findings

• General-purpose video reasoning RL (Video-R1) yields no Jaccard gain over its Qwen2.5-VL-7B backbone on SafeWatch-Bench-Real.
• Extending SFT from 1 to 3 epochs improves Real Jaccard but reduces GenAI Jaccard and cross-dataset VHD11K Jaccard.

Limitations

• Category definitions follow existing benchmarks and may not transfer across cultural, legal, or platform contexts.
• Selecting target FN/FP ratios requires application-specific tuning and validation.
• Intended as an assistive component for human moderators rather than a fully automated system.

Experimental setup

Datasets: SafeWatch-Bench XD-Violence VHD11K-Video

Benchmarks: SafeWatch-Bench-Real XD-Violence SafeWatch-Bench-GenAI VHD11K-Video-Real

Metrics: Jaccard Index Micro-F1 Macro-F1 Binary F1 Precision Recall Error Type Ratio (FN/FP)

Baselines: Qwen2.5-VL-7B Qwen2.5-VL-32B Qwen3-VL-8B Qwen3-VL-235B-A22B InternVL3.5-38B Video-R1 VideoChat-R1 GuardReasoner-VL VAD-R1 SafeWatch-8B GRPO (EM) GRPO (STR)

Builds on: GRPO Tversky Index Video-R1 VAD-R1 SafeWatch-Bench

Closest papers by meaning

• SafeWatch: An Efficient Safety-Policy Following Video Guardrail Model with Transparent Explanations (ICLR 2025 · similarity 0.87)
• Video-SafetyBench: A Benchmark for Safety Evaluation of Video LVLMs (NeurIPS 2025 · similarity 0.86)
• Towards Policy-Adaptive Image Guardrail: Benchmark and Method (CVPR 2026 · similarity 0.86)
• SafeVid: Toward Safety Aligned Video Large Multimodal Models (NeurIPS 2025 · similarity 0.86)
• Diagnosing and Repairing Unsafe Channels in Vision-Language Models via Causal Discovery and Dual-Modal Safety Subspace Projection (CVPR 2026 · similarity 0.85)
• Rethinking Bottlenecks in Safety Fine-Tuning of Vision Language Models (ICLR 2026 · similarity 0.84)