Research Papers
Neural Spectral Capacity Scores Transformers From Specs Alone, No Training Needed
A new NeurIPS 2026 paper introduces Neural Spectral Capacity (NSC), a closed-form scalar that ranks and designs Transformer archit...
Coding Agents Beat Hand-Engineered Planners on Generalized Task and Motion Planning
A new Hugging Face paper finds that coding agents such as Claude Code and Codex can synthesize programs that generalize across tas...
Hugging Face Paper: Hessian-Based RDO Brings Perceptual Metrics In-Loop for Video Codecs
A new Hugging Face paper proposes input-dependent quadratic distortion (IDQD) to approximate full-reference image quality metrics...
ViRDM Cuts the Teacher: Few-Step Causal Video Generation Without DMD's Critic Stack
Hugging Face researchers propose ViRDM, a teacher- and critic-free post-training recipe that matches a precomputed representation...
DeltaWAM Cuts Redundant Video Prediction to Speed Up Bimanual Robot Control
A new world-action model predicts only visual changes rather than dense future frames, boosting success rates on RoboTwin while cu...
Hugging Face Paper Tests RLCD Model Jev as Zero-Shot Detector of AI Alignment Failures
A new Hugging Face paper introduces RLCDAlignBench, benchmarking the RLCD-trained model Jev across ten alignment failure types and...
Hugging Face Paper Proposes a Metric for 'Interesting' Math, Trains 27B Model to Guide Theorem Discovery
A new paper defines the intrinsic interestingness of a theorem as the ratio of proof length to statement length, shows it correlat...
Hugging Face Paper Details Rufus-Air, an Open Eight-Stage LLM Post-Training Recipe
A new paper on Hugging Face presents Rufus-Air, an open and reproducible post-training recipe built on GLM-4.5-Air-Base (106B-A12B...
RGBD20K: Hugging Face Paper Unveils 20,000-Pair RGB-D Segmentation Benchmark
A new Hugging Face paper introduces RGBD20K, a large-scale RGB-D semantic segmentation benchmark with 20,000 image pairs and 160 f...
Sparse Autoencoders Encode Parts of Speech as Distributed, Category-Dependent Groups
A new EMNLP 2026 paper finds that part-of-speech information in LLaMA-3-8B is highly recoverable from sparse autoencoder activatio...
WanPE: Hugging Face Paper Unveils 397B Cinematic Prompt Model for Text-to-Video
A new paper introduces WanPE, a 397B-parameter prompt enhancement model trained on 1.05M real-world videos to bring director-level...
Hugging Face Paper Claims Transformers Can Hold Two Thoughts at Once
A new paper posted to Hugging Face argues that Transformer models exhibit a fundamental linearity, allowing them to superpose the...