Research Papers
UniH3 Framework Unifies Homogeneity and Heterogeneity for All-in-One Medical Image Restoration
A new Hugging Face paper introduces UniH3, a framework that leverages shared anatomical homogeneity alongside task heterogeneity t...
HyQuant Cuts LLM Attention to 4-Bit While Keeping Accuracy Nearly Intact
A new hybrid-precision framework from researchers behind HyQuant keeps a small set of attention-critical tokens and a local window...
Image Tokenizers as Visual Languages: New Study Rethinks Multimodal Model Design
A new study builds a controlled autoregressive testbed to analyze how image tokenizer design shapes joint text-image modeling, fin...
MetroLLM-Bench Tests Small Language Models as Transit Kiosk Brains
A new 955-case benchmark evaluates language models as the policy layer of metro kiosks, finding that a 2.6 GB fine-tuned 4B model...
Hugging Face Paper Proposes Negative Self-Distillation to Fix LLM Reasoning Flaws
A new paper introduces Negative Self-Distillation (NSD), a label-free framework that improves LLM reasoning by pushing models away...
World in World: Training-Free Interface Adds Camera and Time Control to Frozen Video World Models
A new Hugging Face paper introduces World in World, a training-free inference-time interface that routes heterogeneous visual evid...
DRG-MAPPO Brings Graph Attention and Dynamic Roles to Cooperative Air Combat
A new hierarchical multi-agent reinforcement learning framework, DRG-MAPPO, combines graph attention with dynamic role assignment...
New Framework Boosts Multilingual Entity Linking for Rare Entities
A training-free vision-language framework that iteratively retrieves and reasons over Wikipedia improves multimodal entity linking...
IdeaAMBIG Benchmark Finds LLMs Struggle to Spot Missing Details in Research Ideas
A new Hugging Face paper introduces IdeaAMBIG, a 660-instance benchmark testing whether research method specifications are complet...
GLIE Compresses Visual Document Retrieval by Regenerating Embeddings on Demand
A new Hugging Face paper introduces Generative Late-Interaction Embeddings (GLIE), which stores only a few vectors per page and re...
Hugging Face Paper Targets 'Verdict-Preserving' Errors in Autoformalization
A new paper formalizes Verdict-Preserving-Unfaithfulness, a failure mode where flawed formal translations still pass solver checks...
Cohere Labs' Tiny Aya L2-Thinker Reasons in 60 Languages Without Per-Language Training
A new 3.35B-parameter model from Cohere Labs achieves over 93% in-language reasoning across 60 languages by optimizing supervised...