New study finds both men and women engage in gold digging equally, challenging gender stereotypes about relationship wealth-seeking.
RIMap-RISC database models microRNA-mRNA interactions using structural data, enabling precise RNA biology studies and therapeutic design.
Despite 70 years of rule changes and copycat entries, Eurovision remains unpredictable, offering a unique data trove for social science research on cultural evolution.
Canada proposes POET mission to hunt Earth-sized exoplanets as NASA's count nears 6,300, with 223 rocky worlds found. Expert quotes highlight potential for life discovery.
Parker Solar Probe data shows solar radio bursts expose hidden magnetic switchbacks, enabling real-time tracking of Sun's magnetic field.
New blood test detects arsenic exposure via DNA marker, predicts chronic disease risk for 200M+ people worldwide.
Study reveals wild chimpanzees build stronger, insulated nests before rainy nights, showing weather anticipation. This suggests advanced cognitive planning in great apes.
Scientists unveil 'living plastic' that degrades in six days without microplastics, using two bacterial strains. Potential to eliminate persistent plastic waste.
MD Anderson's RF-SIRF maps reversed DNA replication forks in single cells, revealing an epigenetic code for replication stress tied to cancer, aging, and treatment response.
UC study finds gang members openly share activities on TikTok, offering law enforcement a new real-time data tool but raising privacy concerns.
A new method called Anthology uses rich backstories to condition LLMs, enabling more representative, consistent, and diverse virtual personas for research.
A 100-vehicle highway trial proves reinforcement learning can smooth traffic, reduce fuel use, and cut emissions for all drivers, not just AVs.
PLAID uses latent diffusion to generate protein sequences and structures simultaneously from folding model latent spaces, enabling controlled design with function and organism prompts, trained on large sequence databases.
Explains prompt injection attacks on LLMs and proposes two fine-tuning defenses, StruQ and SecAlign, which reduce attack success rates dramatically while preserving utility.
PEVA predicts egocentric video from past frames and whole-body action commands, enabling atomic action generation, counterfactual simulation, and long video prediction for embodied AI.
Word2Vec learns word embeddings via sequential PCA-like factorization, with each concept learned one at a time from small initialization.
Explores divide-and-conquer RL as an alternative to temporal difference learning, addressing error accumulation in long-horizon tasks for scalable off-policy methods.
A new information-theoretic framework evaluates imaging systems by measuring mutual information directly from noisy measurements, outperforming traditional metrics and enabling efficient hardware design.
Explains how SPEX and ProxySPEX algorithms identify critical interactions in LLMs using efficient ablation, overcoming exponential complexity in interpretability.
GRASP introduces three innovations to make gradient-based planning with world models robust for long horizons: virtual states, stochastic exploration, and gradient reshaping.