Papers, benchmarks, and breakthroughs from AI labs — 1629 articles
MarkTechPost · May 24
NVIDIA's Gated DeltaNet-2 decouples erase and write operations in linear attention by using separate channel-wise gates, enabling more flexible memory editing without losing information. At 1.3B param
arXiv ML · May 23
HealthCraft is a reinforcement-learning environment that tests LLM safety in emergency medicine through 205 realistic tasks with 2,337 evaluation criteria, revealing that Claude Opus and GPT-5.4 achie
MIRI · May 22
An OpenAI internal model has autonomously disproven a central conjecture in discrete geometry (the Erdős problem from 1946), with the proof verified by prominent mathematicians who call it a genuine b
arXiv ML · May 23
Researchers prove feature ranking cannot simultaneously be faithful, stable, and complete under collinearity — for correlated features, any ranking method must sacrifice one property. They introduce D
Hugging Face · May 23
Nvidia researchers propose Nemotron-Labs, a diffusion-based language model that generates multiple tokens in parallel instead of sequentially, bypassing the memory bottleneck that makes autoregressive
arXiv AI · May 23
Heavy AI users develop weaker logical reasoning skills than light users in controlled experiments, but high-quality AI assistance preserves learning better than low-quality help—suggesting AI's learni
Hugging Face · May 22
A 3-billion-parameter specialized model for structured OCR outperformed every commercial frontier API tested while costing 50 times less, challenging the enterprise AI assumption that larger models ar
Cureus · May 23
Systematic review comparing AI-based mortality prediction models against traditional ICU scoring systems (APACHE, SOFA, SAPS) across multiple studies. Meta-analysis quantifies whether machine learning