mlsanityData to 14 Oct 2026

BlueGlass: A Framework for Composite AI Safety

arXiv 2507.10106 ↗First posted 14 Jul 2025cs.AIcited by 0

Not on the map. Why?

Abstract (arXiv)

As AI systems become increasingly capable and ubiquitous, ensuring the safety of these systems is critical. However, existing safety tools often target different aspects of model safety and cannot provide full assurance in isolation, highlighting a need for integrated and composite methodologies. This paper introduces BlueGlass, a framework designed to facilitate composite AI safety workflows by providing a unified infrastructure enabling the integration and composition of diverse safety tools that operate across model internals and outputs. Furthermore, to demonstrate the utility of this framework, we present three safety-oriented analyses on vision-language models for the task of object detection: (1) distributional evaluation, revealing performance trade-offs and potential failure modes across distributions; (2) probe-based analysis of layer dynamics highlighting shared hierarchical learning via phase transition; and (3) sparse autoencoders identifying interpretable concepts. More broadly, this work contributes foundational infrastructure and findings for building more robust and reliable AI systems.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 39 of its 59 references that are in the map. First posted on arXiv 14 Jul 2025.

Builds on

39 of 59 references.

References · 39

  1. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,744
  2. 2020An Image is Worth 16x16 Words: Transformers for Image Recognition at Scalecited by 30,186
  3. 2023Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detectioncited by 3,332
  4. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  5. 2019MMDetection: Open MMLab Detection Toolbox and Benchmarkcited by 1,385
  6. 2016Understanding intermediate layers using linear classifier probescited by 1,353
  7. 2018BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learningcited by 1,288
  8. 2019LVIS: A Dataset for Large Vocabulary Instance Segmentationcited by 1,217
  9. 2018Sanity Checks for Saliency Mapscited by 1,156
  10. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 1,032
  11. 2022In-context Learning and Induction Headscited by 964
  12. 2015Deep Learning and the Information Bottleneck Principlecited by 802
27 more
  1. 2019Deep Double Descent: Where Bigger Models and More Data Hurtcited by 672
  2. 2019Object Detection in 20 Years: A Surveycited by 531
  3. 2019On Evaluating Adversarial Robustnesscited by 447
  4. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 435
  5. 2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 422
  6. 2023Florence-2: Advancing a Unified Representation for a Variety of Vision Taskscited by 410
  7. 2024A Survey on Vision-Language-Action Models for Embodied AIcited by 279
  8. 2023Physics of Language Models: Part 3.1, Knowledge Storage and Extractioncited by 270
  9. 2013k-Sparse Autoencoderscited by 215
  10. 2024How to use and interpret activation patchingcited by 214
  11. 2024Transcoders Find Interpretable LLM Feature Circuitscited by 200
  12. 2025Learning Multi-Level Features with Matryoshka Sparse Autoencoderscited by 143
  13. 2025Are Sparse Autoencoders Useful? A Case Study in Sparse Probingcited by 115
  14. 2024A Phase Transition in Diffusion Models Reveals the Hierarchical Nature of Datacited by 93
  15. 2023A Comprehensive Survey on Applications of Transformers for Deep Learning Taskscited by 71
  16. 2023How Deep Neural Networks Learn Compositional Data: The Random Hierarchy Modelcited by 71
  17. 2025Detecting Strategic Deception Using Linear Probescited by 68
  18. 2019Understanding Neural Networks via Feature Visualization: A surveycited by 65
  19. 2024Mamba-360: Survey of State Space Models as Transformer Alternative for Long Sequence Modelling: Methods, Applications, and Challengescited by 64
  20. 2018The EuroCity Persons Dataset: A Novel Benchmark for Object Detectioncited by 57
  21. 2024On Evaluating the Durability of Safeguards for Open-Weight LLMscited by 52
  22. 2023FunnyBirds: A Synthetic Vision Dataset for a Part-Based Analysis of Explainable AI Methodscited by 50
  23. 2024Generative Region-Language Pretraining for Open-Ended Object Detectioncited by 29
  24. 2022Grokking phase transitions in learning local rules with gradient descentcited by 28
  25. 2018A Survey of Word Embeddings Evaluation Methodscited by 26
  26. 2025Safety by Measurement: A Systematic Literature Review of AI Safety Evaluation Methodscited by 15
  27. 2024Beyond the Doors of Perception: Vision Transformers Represent Relations Between Objectscited by 13

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025Self-Aware Safety Augmentation: Leveraging Internal Semantic Understanding to Enhance Safety in Vision-Language Models
  2. 2025Rethinking Bottlenecks in Safety Fine-Tuning of Vision Language Models
  3. 2026Reading, Not Manipulating: Leveraging Router Logits for Multimodal Safety in MoE Vision-Language Models
  4. 2025Agentic Moderation: Multi-Agent Design for Safer Vision-Language Models
  5. 2026Concept-Based Dictionary Learning for Inference-Time Safety in Vision Language Action Models
  6. 2026GuardianBench: A Same-Scene Instruction-Contrastive Benchmark for Latent Contextual Risk in Embodied AI
  7. 2025VLSU: Mapping the Limits of Joint Multimodal Understanding for AI Safety
  8. 2026Automated Safety Benchmarking: A Multi-agent Pipeline for LVLMs