The field of AI research is moving towards a greater emphasis on safety and robustness, with a focus on developing methods to mitigate potential risks and ensure that AI systems behave as intended. Recent work has highlighted the importance of evaluating AI systems in a more comprehensive and nuanced way, taking into account factors such as context, uncertainty, and potential biases. Noteworthy papers in this area include the work on Stress Testing Deliberative Alignment for Anti-Scheming Training, which proposes a framework for assessing anti-scheming interventions and demonstrates the effectiveness of deliberative alignment in reducing covert action rates. Another notable paper is Safe-SAIL, which introduces a framework for interpreting sparse autoencoder features in large language models to advance mechanistic understanding in safety domains.
Advances in AI Safety and Robustness
Sources
Stress Testing Deliberative Alignment for Anti-Scheming Training
Safe-SAIL: Towards a Fine-grained Safety Landscape of Large Language Models via Sparse Autoencoder Interpretation Framework
The Illusion of Readiness: Stress Testing Large Frontier Models on Multimodal Medical Benchmarks
The Ranking Blind Spot: Decision Hijacking in LLM-based Text Ranking
On The Reproducibility Limitations of RAG Systems
Steering Multimodal Large Language Models Decoding for Context-Aware Safety
Extracting Conceptual Spaces from LLMs Using Prototype Embeddings
SCORE: A Semantic Evaluation Framework for Generative Document Parsing
Semantic Representation Attack against Aligned Large Language Models
Solving Freshness in RAG: A Simple Recency Prior and the Limits of Heuristic Trend Detection
What Does Your Benchmark Really Measure? A Framework for Robust Inference of AI Capabilities
LatentGuard: Controllable Latent Steering for Robust Refusal of Attacks and Reliable Response Generation
When Judgment Becomes Noise: How Design Failures in LLM Judge Benchmarks Silently Undermine Validity
RAG Security and Privacy: Formalizing the Threat Model and Attack Surface