Detecting Multi-Turn Attacks in LLMs via Activation Probing
Discover how adaptive probing of LLM activations detects multi-turn adversarial attacks, boosting AI security with 93.8% accuracy.
AdvDMD: High-Quality Few-Step Image Generation Method
Discover AdvDMD, a novel method combining adversarial reward and DMD for efficient, high-quality few-step image generation in diffusion models.
PRISM: Boost Multimodal RL with On-policy Distillation
Discover PRISM, a novel method enhancing multimodal reinforcement learning by reducing distributional drift via black-box on-policy distillation.
Do Sparse Autoencoders Effectively Capture Concept Manifolds?
Explore how sparse autoencoders capture concept manifolds and the implications for representation learning and AI interpretability.
DEFault++: Automated Fault Diagnosis for Transformers
DEFault++ offers precise fault detection, categorization, and diagnosis for transformer models, boosting AI reliability and developer accuracy.