Papers, benchmarks, and breakthroughs from AI labs — 715 articles
Latent Space · Apr 2
Moonlake AI is building interactive, multiplayer world models bootstrapped from game engines—enabling indefinite simulation lifetimes and physics-accurate interactions, contrasting sharply with single
arXiv ML · Apr 3
Researchers trained LLMs to forecast supply chain disruptions with calibrated probabilistic outputs, outperforming GPT-5 on accuracy and calibration. The team open-sourced their evaluation dataset to
MarkTechPost · Apr 3
TII's Falcon Perception unifies vision and language in a single 600M-parameter Transformer, replacing the standard encoder-decoder pipeline with early fusion from the first layer. The model uses hybri
arXiv ML · Apr 3
Sven, a new optimizer using truncated SVD of the loss Jacobian, converges faster than Adam on regression tasks with computational overhead only linear in the number of retained singular values, compar
Anthropic · Apr 2
Anthropic researchers discovered that Claude organizes emotion concepts along interpretable dimensions (valence, arousal, agency) in its internal representations, similar to how neuroscientists model
arXiv ML · Apr 3
Tutorial explaining how Bayesian optimization—using surrogate models and acquisition functions to balance exploration and exploitation—systematizes the hypothesis-experiment-refine cycle in scientific
Hugging Face · Apr 1
Hugging Face unveiled Holo3, an open-weight computer-use agent that scores 78.85% on OSWorld-Verified (new SOTA), using only 10B active parameters out of 122B total—matching or beating closed models l