6arXiv cs.CL (Computation and Language)·
Shieldstral: 3B multimodal safety classifier matching models 7× its size
Shieldstral is a 3B-parameter policy-adaptive multimodal safety classifier that matches or outperforms models nearly 7× its size on text safety benchmarks and claims state-of-the-art on multimodal safety classification. The model frames content moderation as a binary yes/no question-answering task, allowing heterogeneous safety datasets with divergent taxonomies to be unified under a single training framework. The authors describe a data construction pipeline covering ~54.1M samples and a fine-grained evaluation set for policy adaptability. The efficiency-at-scale result is notable for practitioners deploying content moderation in resource-constrained settings.
Related guides (3)
Related events (8)
Meta releases Llama Guard 4 12B multimodal safety classifier on Hugging Face
Meta released Llama Guard 4 12B, a multimodal (image-text-to-text) safety classification model built on the Llama 4 architecture, published to Hugging Face. The model is designed for conversational safety filtering and supports both text and image inputs. With 143K downloads and 102 likes shortly after release, it is seeing meaningful early adoption.
Meta releases Llama Guard 3 1B safety classifier on Hugging Face
Meta released Llama Guard 3 1B, a compact 1-billion-parameter text-generation model designed for content safety classification, published on Hugging Face. The model is part of the Llama Guard 3 family and supports multiple languages including English, German, and French. Its small size makes it suitable for lightweight safety filtering in production deployments.
Qwen3Guard: Real-time Safety Guardrail Model for Token Stream Classification
Alibaba's Qwen team has released Qwen3Guard, the first dedicated safety guardrail model in the Qwen family, built on Qwen3 foundation models and fine-tuned for safety classification. The model performs real-time safety detection on both prompts and responses, providing risk levels and categorized classifications for content moderation. Qwen3Guard claims state-of-the-art performance on major safety benchmarks across English, Chinese, and multilingual settings.
Meta releases Llama Guard 3 11B Vision for multimodal content safety classification
Meta released Llama Guard 3 11B Vision on Hugging Face, a multimodal safety classifier supporting image-text-to-text inputs built on the Llama 3 architecture. The model extends the Llama Guard safety classification family to handle visual content alongside text. This is relevant to AI safety tooling for multimodal deployments.
SafeSteer: Localized On-Policy Distillation for Efficient Safety Alignment
SafeSteer proposes a safety alignment method that targets only 'safety tokens' in the output distribution rather than applying global fine-tuning, arguing that safety features are inherently sparse. It constructs a safety teacher via activation steering, then restricts a reverse KL penalty to selected safety tokens during training. The approach achieves strong safety performance across seven benchmarks with minimal capability degradation, requiring only 100 harmful samples—less than 1% of data used by prior baselines.
NVIDIA releases Nemotron 3.5 Content Safety, a customizable multimodal safety model for enterprise AI
NVIDIA has released Nemotron 3.5 Content Safety, a multimodal safety model designed for enterprise AI deployments with customization capabilities for global use cases. The model is announced via the Hugging Face blog, targeting content moderation and safety classification across modalities. This is relevant to the growing enterprise demand for controllable, deployable safety layers on top of foundation models.
PaSBench-Video: A Streaming Video Benchmark for Proactive Safety Warning in MLLMs
PaSBench-Video is a 740-video benchmark designed to evaluate whether multimodal large language models can issue timely, accurate safety warnings during the window between a visible danger sign and an accident. Videos span four domains (driving, healthcare, daily life, industrial production) and are annotated with frame-level risk onset and accident boundaries, requiring causal temporal reasoning rather than static scene classification. Testing 13 MLLMs reveals no model exceeds 20% on the strictest metric, with recall strongly coupled to false-positive rate (Pearson r=0.64), indicating models rely on scene-level activity cues rather than genuine hazard reasoning. Performance varies sharply by domain, with driving being particularly problematic due to visual similarity between routine and hazardous scenes.
Systematic comparison of encoder vs. decoder safety judges for LLM adversarial evaluation
A new arXiv preprint evaluates whether fine-tuned encoder classifiers from the ModernBERT family (ModernBERT and Ettin) can replace LLM-based safety judges for detecting harmful outputs in user-model conversations. The study benchmarks encoders against rule-based methods, fine-tuned LLM classifiers, and LLM judges including LlamaGuard 3/4, ShieldGemma, StrongReject, and Claude-as-a-judge across multiple adversarial attack types. Results are reported on F1, false negative rate, and precision-recall, with breakdowns by attack technique, providing practical guidance on cost-latency tradeoffs for production safety pipelines.