Researchers are uncovering new attack surfaces: visual jailbreaks, cumulative dialogue risks, watermark-induced hallucinations in medicine, and the structural limits of Transformers. These findings question the reliability of current LLMs in critical applications.
Researchers find that multimodal LLMs robustly understand content regardless of visual style, yet their safety mechanisms can be easily bypassed by specific stylistic triggers. The new ASO approach automates adversarial image creation by optimizing s...
A new technique exploits the Mixture of Experts architecture to identify and correct hallucinations in LLMs at inference time, without altering the model. The study reveals that in higher MoE layers, different expert groups activate in opposite ways ...
A new study shows that watermarks on medical LLMs induce hallucinations, lexical corruption, and diagnostic errors. Standard benchmarks fail to catch these failures, creating a false sense of safety. The issue is critical for on-premise healthcare de...
A formal study shows that pure Transformers, bounded by class TC⁰, cannot learn structural generalization, which is NC¹-complete. Neuro-symbolic systems bypass the obstacle by injecting the semantic component, but benchmarks fail to separate learned ...
Today's guardrails evaluate each prompt-response pair in isolation, overlooking risks that emerge only over multi-turn dialogues. A new framework tracks semantic drift and information accumulation, with direct implications for on-premise deployment a...
A new operator, HyenaND, promises to replace attention on multidimensional data using input-dependent long convolutions. Subquadratic O(L log L) scaling and an optimized CUDA implementation deliver real speedups on genomics, medical imaging, and PDE ...
