Research Papers
Hugging Face Researchers Unveil 'Compile by Training' to Turn Natural Language into Local Neural Functions
A new technique called 'compile by training' converts natural-language specifications into reusable neural functions by distilling...
Hugging Face Unveils Puffin-World: A Unified Model for 3D World Generation and Exploration
Puffin-World is a unified multimodal framework that jointly models physics, geometry, and appearance for physically consistent 3D...
One Query Enough? Hugging Face Study Rethinks Data Needs in On-Policy Distillation
A new Hugging Face paper shows that on-policy distillation (OPD) can recover most of its full-data performance using just a single...
Random Attention: KV Cache Eviction Without Scoring Matches Selective Methods
A new paper from Salesforce AI Research shows that randomly evicting cached tokens in LLM reasoning tasks performs as well as soph...
Gated DeltaNet Survives 4-Bit Quantization: New Recipe Quantizes All Layers of Hybrid 27B LLM
A new study shows that fully quantizing hybrid LLMs, including recurrent Gated DeltaNet layers, to 4-bit NVFP4 preserves accuracy...
LatentPress: Compressing AI Context into Direct-Read Memory Tokens
Hugging Face researchers introduce LatentPress, a method that compresses conversational and document context into continuous memor...
Hugging Face Unveils LLaDA-Image: Open-Source 6B Model for Photorealistic Generation and Editing
LLaDA-Image pairs a 6B diffusion transformer with a frozen vision-language module, using image-only pre-training and the Muon opti...
NVIDIA Nemotron Models Outperform Top Humans at IOI 2026 Coding Competition
A new post-training pipeline from NVIDIA, combining curated problems, synthetic reasoning, SFT, and RL, enables Nemotron models to...
New Distillation Method Tackles Diversity Bottleneck in Reasoning Models
Researchers propose IDA-OPD, a distillation technique that improves diversity transfer in reasoning models by selectively preservi...
Cliff: New Reward Shaping Method Boosts LLM Reasoning by Detecting First Mistakes
Researchers propose Cliff, a reward shaping strategy that uses an off-the-shelf LLM to detect the first reasoning error, improving...
Hugging Face Unveils VibeVoice-ASR-Streaming: Real-Time Speaker-Attributed Speech Recognition
Researchers present VibeVoice-ASR-Streaming, an LLM-based end-to-end model that unifies streaming speech recognition and speaker d...
SimLoss: Single-Pass Image Captioning Matches Multi-Stage Quality at 20x Lower Latency
A new Hugging Face paper introduces SimLoss, an embedding-space contrastive objective that enables single-pass fine-grained image...