Research Papers
EviRank: Training-Free Multimodal Image Re-ranking via Structured Evidence
Hugging Face researchers introduce EviRank, a training-free method that reformulates multimodal image re-ranking as semantic const...
Graph Engineering: A New Paradigm for Coordinating LLM Agents into System Intelligence
A new survey from Hugging Face introduces Graph Engineering, a paradigm that uses dynamic graph structures to organize multi-agent...
New Benchmark Reveals Omni-LLMs Struggle as Real-Time Video Assistants
Hugging Face researchers introduce OmniAssistBench, a benchmark for evaluating omni-modal LLMs as interactive video assistants. Re...
Hugging Face Proposes Compute-Efficient Hyperparameter Transfer for Large-Scale MoE Models
A new framework from Hugging Face predicts optimal learning rates for large Mixture-of-Experts models by scaling across widths and...
FlowEvo: Self-Evolving Agents Co-Develop Workflows and Skills at Inference Time
FlowEvo, a training-free framework from Hugging Face, enables LLM agents to co-evolve reusable skills and workflows during inferen...
NAPE: A Minimalist Causal Transformer for Scalable Audio Self-Supervised Learning
Researchers introduce NAPE, a self-supervised audio learning framework that uses causal Transformers to predict next spectrogram p...
PolicyGuide: A New Framework to Enforce Policy Compliance in LLM Agents Across Entire Workflows
Hugging Face researchers introduce PolicyGuide, a framework that compiles domain policies into workflow graphs and uses a proactiv...
Hierarchical Self-Improvement: Evolving Agent Harnesses for Frozen LLMs
A new framework from HKUST lets a frozen LLM evolve its own task-specific execution harness and evolution strategy, yielding signi...
New NARU Benchmark Tests AI's Grasp of Japanese Long-Form Video Narratives
Researchers introduce NARU, a benchmark with 1,481 questions across 155 Japanese videos (146.8 hours) to evaluate narrative evolut...
TinyCast: A 146K-Parameter Zero-Shot Forecaster That Runs on Embedded Devices
Hugging Face researchers introduce TinyCast, an attention-free zero-shot forecaster with only 146,505 parameters that computes per...
New Study Quantifies Benchmark Optimization in ASR Models, Revealing Inflated Scores
A new paper from Hume AI introduces a methodology to quantify benchmark optimization in ASR models, showing that top-scoring model...
EXIMO: A Three-Stage Method to Efficiently Fine-Tune VLA Robot Policies
Researchers propose EXIMO, a novel algorithm that combines VLM-guided exploration, imitation learning, and residual off-policy RL...