Research Papers
TempCloze Benchmark Tests Whether Video-LLMs Can Find the Missing Middle
Hugging Face paper introduces TempCloze, a video cloze benchmark that asks Video-LLMs to pick the true missing middle segment from...
Hugging Face Releases Open Recipe for IMO Gold With Nemotron
A new open pipeline built on post-trained Nemotron 3 Ultra checkpoints scored 30 of 42 points at IMO 2026, reaching the gold-medal...
FreeFlow: A Bias-Free Hierarchical Transformer Sets New Optical Flow Records
Hugging Face paper introduces FreeFlow, a hierarchical transformer that drops flow-specific inductive biases such as correlation v...
Hugging Face Paper Unveils MaP-WAM, a Memory-as-Plans Framework for Non-Markovian Robot Manipulation
MaP-WAM separates memory-grounded planning from plan-conditioned execution, using episodic segment records and progress-calibrated...
X-AuT Cuts Speech-LLM Audio Encoders While Improving Accuracy
A new progressive compression framework from XPENG-AI prunes audio-encoder layers in speech LLMs and recovers accuracy through cro...
Mi-Ripple Targets 'Digital Ripple' Artifacts Left by Iterative AI Image Editing
A new Hugging Face paper introduces Mi-Ripple, a diagnosis-guided restoration workflow that separates periodic lattice artifacts f...
CARDEA Brings Auditable Spatial Reasoning to Coronary Angiography AI
A new vision-language model, CARDEA, uses Chain-of-Box reasoning and reinforcement learning with verifiable rewards to expose the...
Hugging Face Paper Unveils RCWM: Recursive Code World Models for 3D Scene Reconstruction
A new research paper introduces Recursive Code World Models (RCWM), a framework that reconstructs complex 3D worlds as executable...
SpatialBlock: Synthetic Block-Stacking Boosts 3D Spatial Reasoning in LVLMs
Hugging Face paper introduces SpatialBlock-15k, a synthetic dataset of 15,000 block-stacking problems that improves 3D spatial rea...
EvoSafeHarness Evolves Custom Safety Harnesses for LLM Agents
A new framework jointly searches natural-language policies and executable enforcement code to build safety harnesses tailored to a...
SenseNova-U1.5 Debuts as 8B Encoder-Free, VAE-Free Unified Visual Model
SenseNova-U1.5 is an 8B native unified multimodal model that handles visual understanding, reasoning, and generation without encod...
Hugging Face Paper: 8.9B Latent-Space Model Beats OLMo-3-7B With Half the Tokens
NCP-ArchPreview jointly trains next-token and next-concept prediction, reaching OLMo-3-7B's final pretraining loss using only 51.3...