Research Papers
CW-BASS v2: Saturation-Aware Pseudo-Label Selection for Semi-Supervised Segmentation
A new method, CW-BASS v2, adapts pseudo-label selection to the saturated confidence of foundation-model teachers, improving semi-s...
Inaudible Low-Frequency Attacks Can Cripple Audio-Language Models, New Study Warns
Researchers propose a black-box red-teaming method using inaudible low-frequency waveforms to expose vulnerabilities in audio-lang...
AI Agent Refactors 717k-Line Codebase Without Human Review: A Specification-First Case Study
A new case study shows an AI coding agent successfully dismantling a core architectural invariant across 189 files in a 717k-line...
New RL Method Teaches LLMs When to Stop and Refuse Futile Reasoning
Researchers introduce CaRL, a reinforcement learning approach that aligns LLM behavior with capability boundaries, reducing futile...
SKILLER: RL Framework Cuts Inference Costs for Small Language Models
Hugging Face researchers introduce SKILLER, a natural-language-driven reinforcement learning framework that automatically generate...
OmniScientist: An Omni-Modal AI Scientist That Reasons Directly from Raw Evidence
Hugging Face researchers introduce OmniScientist, an end-to-end AI scientist that autonomously conducts multidisciplinary research...
Hugging Face's AVA-Encoder: Agent-Native Video Representation via Knowledge Graphs
AVA-Encoder learns structured video representations via agentic auto-encoding using knowledge graphs, enabling cinematic video gen...
H2R-Bench: New Benchmark Tests AI's Human-to-Robot Video Generation
Hugging Face researchers introduce H2R-Bench, a benchmark for evaluating video generation models that transform human manipulation...
AutoPrune: LLM-Designed Visual Token Pruning Cuts 94.4% Tokens with 99% Performance
Hugging Face researchers introduce AutoPrune, a training-free framework that uses large language models to automatically design vi...
Full-Bandwidth Transformers: Latent Feedback Boosts Reasoning and Efficiency
A new Hugging Face paper introduces full-bandwidth transformers, which feed the top-layer hidden state back into the model via a g...
HPSE: Hybrid-Policy Self-Editing Boosts Composable Knowledge Editing in LLMs
Researchers propose HPSE, a method that improves unstructured knowledge editing by distilling from hybrid rollouts that insert mis...
Instruction Tuning Alters Confidence and Reduces Rationale Diversity Without Improving Calibration
A new study finds that instruction tuning consistently changes model confidence and reduces cross-rationale diversity, yet does no...