Related papers: Shieldstral

Shieldstral

URL: http://arxiv.org/abs/2607.25857v2
Date: Tue, 04 Aug 2026 18:10:46 GMT
Title: Shieldstral
Abstract summary: Shieldstral is a policy-adaptive multimodal safety classifier that matches or outperforms models nearly 7$times$ its size on text safety benchmarks.<n>We present a data construction recipe, covering curation and generation of approximately 54.1M samples and a fine-grained evaluation set to evaluate policy adaptability.
Score: 183.97744446582863
License: http://creativecommons.org/licenses/by/4.0/
Abstract: We introduce Shieldstral, a 3B-parameter policy-adaptive multimodal safety classifier that matches or outperforms models nearly 7$\times$ its size on text safety benchmarks and sets a new state of the art on multimodal safety classification. Shieldstral formulates content moderation as a binary question-answering task. This simple formulation unifies diverse moderation tasks into a single yes/no problem, enabling heterogeneous safety datasets with divergent taxonomies to be consolidated under one training framework. We present the data construction recipe, covering curation and generation of approximately 54.1M samples and a fine-grained evaluation set to evaluate policy adaptability. Together, these enable a small adaptive model to match or outperform much larger models.

Related papers

SafeAtlas-VL: Beyond Binary Multimodal Safety with Large-Scale Data and Guard Models [73.04773649437375]
Multimodal safety moderation requires distinguishing risks arising from visual content, user intent, and assistant behavior.<n>We introduce SafeAtlas-VL, a dataset of 1.5M training instances that places image-, request-, and response-level judgments on a five-level ordered scale.<n>We also construct SafeAtlas-Bench, a set of 5,000 instances for evaluating five-level predictions and continuous risk scores.
arXiv Detail & Related papers (2026-08-29T07:08:01Z)
Selective Safety Steering via Value-Filtered Decoding [54.87935112120107]
Large language models (LLMs) are trained to align with human values, but their generations may still violate safety constraints.<n>Existing decoding-time steering methods often intervene unnecessarily, modifying generations that would have been safe under the base model.<n>We propose a new test-time steering method designed to reduce such unnecessary interventions while improving the safety of unsafe responses.
arXiv Detail & Related papers (2026-05-14T12:13:08Z)
Escaping the Diversity Trap in Robotic Manipulation via Anchor-Centric Adaptation [50.23374353859762]
"Maximizing coverage" by collecting diverse, single-shot demonstrations can be self-defeating due to non-vanishing estimation noise.<n>We formalize this phenomenon as a Coverage--Density Trade-off.<n>We propose Anchor-Centric Adaptation (ACA), a two-stage framework that first stabilizes a policy skeleton through repeated demonstrations at core anchors, then expands coverage to high-risk boundaries via teacher-forced error mining and constrained residual updates.
arXiv Detail & Related papers (2026-05-08T07:35:24Z)
EnchTable: Unified Safety Alignment Transfer in Fine-tuned Large Language Models [23.11474404054016]
Many machine learning models are fine-tuned from large language models (LLMs) to achieve high performance in specialized domains like code generation, biomedical analysis, and mathematical problem solving.<n>We introduce EnchTable, a novel framework designed to transfer and maintain safety alignment in downstream LLMs without requiring extensive retraining.
arXiv Detail & Related papers (2025-11-13T02:26:59Z)
Reimagining Safety Alignment with An Image [49.33281424100804]
Large language models (LLMs) excel in diverse applications but face dual challenges: generating harmful content under jailbreak attacks and over-refusal of benign queries.<n>We propose Magic Image, an optimization-driven visual prompt framework that enhances security while reducing over-refusal.
arXiv Detail & Related papers (2025-11-01T11:27:07Z)
Benchmarking Unified Face Attack Detection via Hierarchical Prompt Tuning [58.16354555208417]
PAD and FFD are proposed to protect face data from physical media-based Presentation Attacks and digital editing-based DeepFakes, respectively.<n>The lack of a Unified Face Attack Detection model to simultaneously handle attacks in these two categories is mainly attributed to two factors.<n>We present a novel Visual-Language Model-based Hierarchical Prompt Tuning Framework that adaptively explores multiple classification criteria from different semantic spaces.
arXiv Detail & Related papers (2025-05-19T16:35:45Z)
Aegis2.0: A Diverse AI Safety Dataset and Risks Taxonomy for Alignment of LLM Guardrails [4.697160328460634]
Large Language Models (LLMs) and generative AI become increasingly widespread.<n>There is a clear lack of high-quality, human-annotated datasets that address the full spectrum of LLM-related safety risks.<n>We propose a comprehensive and adaptable taxonomy for categorizing safety risks, structured into 12 top-level hazard categories with an extension to 9 fine-grained subcategories.
arXiv Detail & Related papers (2025-01-15T18:37:08Z)
Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements [46.79887158348167]
The current paradigm for safety alignment of large language models (LLMs) follows a one-size-fits-all approach.<n>We propose Controllable Safety Alignment (CoSA), a framework designed to adapt models to diverse safety requirements without re-training.
arXiv Detail & Related papers (2024-10-11T16:38:01Z)

This list is automatically generated from the titles and abstracts of the papers in this site.