Research Papers
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...
LycheeMemory V2: Segment-Level Memory Consolidation Cuts LLM Agent Costs by 86%
Hugging Face researchers introduce LycheeMemory V2, a long-term memory framework that batches interactions into semantic segments,...
Hugging Face Unveils Maglev: A Recurrent Transformer with Sliding Memory for Efficient Long-Context Modeling
Maglev introduces a recurrent Transformer architecture with fixed-size memory that generalizes sliding-window attention, enabling...
Context-Matched Distillation Aligns Teacher Supervision for Autoregressive Video Generation
A new method, Context-Matched Distillation (CMD), aligns teacher supervision with causal generation context in few-step autoregres...
Gambit: Thought-Level Beam Search Boosts Reasoning Efficiency by Up to 68.5%
Hugging Face researchers introduce Gambit, an inference algorithm that uses thought-level beam search to dynamically allocate comp...
LiveAnimate: Real-Time Long-Form Human Animation with 14B Diffusion Transformer
Hugging Face researchers introduce LiveAnimate, the first system to combine real-time streaming with stable long-form human animat...
UniSwap: First Streaming Framework for Joint Audio-Visual Identity Swap in Talking Videos
Hugging Face researchers introduce UniSwap, a unified streaming audio-visual diffusion transformer that simultaneously transfers a...
Massive Activations in Hybrid Linear Attention LLMs: New Study Reveals Pre-Attention Spikes and Inter-Spike Plateaus
A new study from Hugging Face provides the first systematic analysis of massive activations in hybrid linear-attention LLMs, uncov...
PlayWorld Benchmark Puts World Models to the Test with Agent Players
Hugging Face researchers introduce PlayWorld, a benchmark that uses multimodal agent players to evaluate interactive video world m...
Rhetoric Can Hack AI Peer Review: New Study Reveals Structured Sensitivity
A new study from Hugging Face shows that rhetorical framing biases AI review scores in structured ways, with evidence framing and...
Hugging Face Unveils Evoke: An Interactive World Model with Persistent Memory for Endless Video Generation
Evoke, a new interactive world model from Hugging Face, uses external persistent memory and a redesigned long-horizon teacher to e...