Archive, page 2
A snapshot of arXiv metadata, version history, submission files and rendered documents covering 3,148,796 papers, with a PDF for 99.47% of them, plus SHA-256 digests.
A 1K to 10K-row Apache 2.0 dataset of doctor-patient ChatML conversations covering every human disease on a list, generated by Claude Opus 5.5, one agent per disease. Sequel to an Opus 4.8 version.
An Apache 2.0 benchmark where a model reads one unstructured state and answers five typed questions, each as a probability distribution. Schema follows TypeSafe AI primitives so it runs against any matching API.
Twelve separate frozen data configs from the Open-Jev project, turning a state and question into a yes/no probability, choice distribution, label probabilities or numeric decision. Includes source code and manifests.
A Parquet dataset of 64,964 normalized Indian public information document texts with source titles, URLs, extraction quality and topic labels, generated October 2026.
An MIT-licensed dataset of over 5,500 RL tasks in code and data science for hill-climbing small models, with 144 held-out tasks. A 2B model was trained on the tasks, with a curriculum run compared to shuffled.
Hugging Face blog post on how scheduling works for GPU clusters, explaining approaches to allocating compute across training and inference jobs.
A paper introducing Memento 3, which lets frozen LLM agents keep learning by revising explicit world models from new observations. It has 28 upvotes on Hugging Face Papers.
A paper on dense correspondence matching that drops simple motion and rigid-geometry priors, aimed at image editing and reference-guided generation where identity must be kept. It has 34 upvotes.
A technical report on the MiMo-V2.6 omni-modal model family, which scales reinforcement learning compute after multimodal mid-training. It has 65 upvotes on Hugging Face Papers.
A book on large language model foundations, structured into six chapters on pre-training, generative models, prompting, alignment, inference and reasoning. Aimed at students and practitioners. It has 38 upvotes.
A research framework that turns any static 3D Gaussian Splatting scene into a 3D cinemagraph with looping motion from any viewpoint. A vision-language model guides a video model to create the motion. 37 upvotes.
A paper on egocentric world models that predict first-person views for multiple agents acting in a shared environment, going beyond coarse actions like locomotion. It has 47 upvotes.
A paper introducing OneSearch-VL, an agent that handles deep research over single images, multiple images and video with visual grounding, retrieval and fact composition. It has 20 upvotes.
A paper introducing TestPrism, with 300 test tasks from 17 sources and 3,000 candidate implementations split between valid and invalid. It argues single-reference evaluation overstates test quality. 30 upvotes.