Research Papers
ElephantBench: Probing LLMs' Blind Spots in Long-Tail Knowledge
A new benchmark reveals that even the strongest LLMs fail to recall multiple conflicting accounts of long-tail facts, with the bes...
Parallel Tube Decoding Cuts Video Grounding Latency 79x
Hugging Face researchers introduce Parallel Tube Decoding (PTD), a generative method that decodes spatial and temporal video groun...
Hugging Face Paper Proposes Fast Weight Attention for Continual Learning
A new research paper on Hugging Face introduces Fast Weight Attention, a framework that unifies recurrent fast-weight memories and...
StarHarness: Evolving Agent Harnesses Boosts Enterprise Performance by 20–35 Points
New research from Hugging Face introduces StarHarness, a framework that evolves agent harnesses around fixed model weights, improv...
LayerRecall: A New Memory Router Boosts Long-Video Consistency in Diffusion Models
Hugging Face researchers introduce LayerRecall, a state-conditioned memory router that selectively injects long-range historical m...
Intention Distillation: Teaching Robot Action Decoders the 'Why' Behind Behavior
Hugging Face researchers introduce Intention Distillation (INDI), a method that distills behavior-level intent from a frozen teach...
Puro-2B: Training a 2B LLM for Under $5,090 on Consumer GPUs
A new open-source recipe trains 2B-parameter LLMs on RTX 5090 GPUs for under $7K, approaching Qwen2.5-1.5B performance and derivin...
Survey Defines Agentic Artifact Creation: 230 Systems Analyzed, Principles Proposed
A new survey from Hugging Face defines 'agentic artifact creation' as stateful, feedback-driven AI construction of deliverables, a...
DART-SD: A New Framework to Fix Multi-Turn Tool-Calling Agents by Targeting Critical Breakpoints
Hugging Face researchers introduce DART-SD, a topology-aware self-distillation framework that models tool-use trajectories as diam...
LoopArena: New Benchmark Tests AI Models as Runtime Controllers for Coding Agents
Hugging Face researchers introduce LoopArena, a benchmark that evaluates how well a controller model guides a separate coding agen...
Hugging Face Unveils StepGuard: Step-Level Guardrails for Safer AI Agents
StepGuard, a new step-level guard model from Hugging Face, audits agent actions before execution, reducing attack success rates by...
Hugging Face Researchers Unveil ABot-Recon for Stable Long-Horizon 3D Reconstruction
ABot-Recon, a new streaming 3D reconstruction model from Hugging Face, achieves stable long-horizon performance using only local t...