Research Papers
Google DeepMind Unveils AlphaGenome Atlas: Predictive Map of 9 Billion Human DNA Variants
Google DeepMind introduces AlphaGenome Atlas, a comprehensive platform with molecular predictions for all 9 billion possible singl...
Split-LLM Training Leaks Real Data Through Zero-Gradient Patterns, Study Finds
A systems-security case study shows that a split-LLM training setup passed its privacy checks while the returned gradient revealed...
SimpleMemVLA Drops Dedicated Memory Modules for Raw Video History in Robot Policies
A new vision-language-action model feeds intact, timestamped video history straight into a pretrained VLM backbone, letting hidden...
Hugging Face Paper Proposes Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal
A new Hugging Face paper introduces narrow-boundary safety alignment, using self-generated refusal data and harmful-benign boundar...
Hugging Face Researchers Unveil Conformal Relevance Framework for NLP Tasks
A new framework from Hugging Face automates score function design for conformal prediction using in-context learning and ensemblin...
On-Policy Self-Distillation: A Critical Review Maps the Three Levers of Reasoning Collapse
A new critical review examines on-policy self-distillation (OPSD) for mathematical reasoning, framing reasoning collapse as a symp...
EmbodiedSkills: A Unified Framework for Closed-Loop VLA Agents
Hugging Face researchers propose EmbodiedSkills, a framework that verifies robot skill executions via a fixed interface, enabling...
Hugging Face Paper Introduces Causal Foundation Models for In-Context Causal Inference
A new paper from Hugging Face presents Causal Foundation Models (CFMs), pretrained networks that estimate causal effects on new da...
Hugging Face Unveils ENEAS: A Text-Promptable Method for Robust Instance Tracking and Semantic Discovery
ENEAS unifies text-prompted instance tracking and open-concept semantic discovery, addressing temporal hallucinations, spatial fra...
Hugging Face Researchers Unveil Uno: Lossless Diffusion Speedups for LLMs
A new class of diffusion-augmented LLMs, named Uno, enables parallel token sampling to accelerate inference by up to 3x without qu...
Hugging Face Researchers Introduce RevPropBench to Test LLMs' Revision Propagation in Conversational Artifacts
A new benchmark, RevPropBench, evaluates how well LLMs propagate local revisions across dependent elements in conversationally gen...
Hugging Face Unveils FlowBalance: Verifier-Grounded Self-Improvement for Reasoning Models
FlowBalance enhances reasoning models by calibrating self-guidance with verifier feedback, improving performance and stability ove...