Jailbreaking Frontier AI Models via Intention Deception
Explore how multi-turn intention deception exploits vulnerabilities in frontier AI models like GPT-5, revealing critical safety risks and para-jailbreaking...
Discovering LLM Personas via Bridging Inference Analysis
Explore how bridging inference reveals stable, coherent personas in large language models, enhancing AI dialogue understanding and interaction.
FreeScale: Efficient Distributed Training for Recommendation Models
Discover FreeScale, a method that reduces scaling costs and boosts efficiency in distributed training of sequence recommendation models.
Improving Verbal Confidence in Gemma 3 4B LLMs
Explore how confidence-conditioned fine-tuning enhances verbal accuracy in Gemma 3 4B, revealing key lessons on label entropy and output regularization.
QEVA: Reference-Free Metric for Narrative Video Summarization
Discover QEVA, a novel reference-free metric using multimodal QA to evaluate narrative video summaries on coverage, factuality, and chronology.