LLM Detection
A new research paper introduces Latent Prompt Restoration, a detection method that identifies LLM-generated text by exploiting the structural traces prompts leave behind in model outputs, offering a fresh angle on synthetic text authenticity.
LLM Safety
New research tackles the 'hidden puppet master' problem: predicting how manipulative LLM dialogues change human beliefs. The work offers a framework for detecting persuasion and covert influence in conversational AI systems.
Sakana AI
Sakana AI's error diffusion method trains Dale-compliant dual-stream networks without backpropagation, hitting 96.7% on MNIST and 61.7% on CIFAR-10, pointing toward more biologically plausible learning.
data provenance
OriginBlame introduces a technique for tracing data provenance at both the record and token level in AI training datasets, offering new tools for accountability, attribution, and authenticity in the era of generative models.
deepfake detection
Researchers at Brazil's Unicamp have developed a deepfake detection approach that generalizes to synthetic media created with techniques the system has never seen, tackling one of the toughest problems in digital authenticity.
AI Safety
New research proposes verifiable steering and refusal mechanisms for autonomous AI scientists, addressing when discovery agents should halt experiments based on safety, validity, or resource constraints in scientific workflows.
AI Research
A new position paper argues that a true science of AI requires studying training dynamics directly rather than relying on post-hoc fixes, with implications for understanding generative models and synthetic media systems.
Andrej Karpathy
Andrej Karpathy compresses 33 years of neural network evolution into a 90-second retrospective, tracing the architectural lineage from early MLPs to today's transformer-based generative models powering synthetic media.
model collapse
A new arXiv paper frames model collapse from synthetic data contamination as an epidemiological problem, applying bilayer SIR dynamics to model how AI-generated content spreads through training corpora and degrades future models.
LLM Evaluation
New research challenges the assumption that agreement between LLM judges signals human alignment. The geometry of LLM evaluation reveals systematic biases that affect how synthetic content and AI outputs are assessed.
LLM watermarking
New research proposes a linguistics-aware watermarking method for LLMs that embeds detectable signals into generated text without distorting output quality, advancing AI content authentication.
audio deepfakes
AAAI research benchmarks state-of-the-art audio deepfake detectors against real social media content, exposing a stark gap between lab accuracy and field performance on TikTok, YouTube, and X.