Research Papers
EM²Mem: Event-Centric Memory Framework Boosts Long-Video QA
A new Hugging Face paper introduces EM²Mem, an event-centric multimodal memory framework that binds evidence to event anchors, imp...
E-Commerce Bench: New Benchmark Tests LLM Agents on Year-Long Business Operations
Hugging Face researchers introduce E-Commerce Bench, the first open-source benchmark simulating a year-long e-commerce business fo...
Harness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement
A new framework called Harness-of-Harness (HoH) enables coding agents to iteratively improve software through structured loops, ac...
DiagEvo: Self-Evolving LLMs Learn from Their Own Mistakes
New research introduces DiagEvo, a method that lets language models improve by diagnosing their own failure history, outperforming...
AgentJudgeBench: LLM Judges Hit Structural Ceiling on Complex Agentic Workflows
A new benchmark reveals that LLM judges degrade sharply with task difficulty on agentic tool-calling workflows, converging to a 77...
Hugging Face Unveils Safin-1: A Memory-Native Architecture for 'Safety from Within'
Safin-1 embeds safety as an evolving internal state via memory routing, moving beyond external guardrails to make safety intrinsic...
Study Reveals How Understanding and Generation Truly Synergize in Unified Multimodal Models
New research from Hugging Face explores when visual understanding and generation reinforce each other in unified multimodal models...
DroneCATS: Small Open Models Fly Well but Fail at Protocol, New Benchmark Shows
A new benchmark, DroneCATS, evaluates multimodal LLMs as drone controllers, revealing that small open models often navigate better...
Hugging Face: Self-Hosted LLM with GRPO Experts Merged via SLERP Beats 7x Larger Model
A new research paper from Hugging Face shows that a smaller self-hosted LLM, trained with separate GRPO experts merged via SLERP,...
Qwen-Drive-1.0: A Vision-Language Foundation Model for Autonomous Driving
Hugging Face researchers introduce Qwen-Drive-1.0, a vision-language foundation model that unifies 3D perception, visual question...
H3-World: Turning Language Understanding into Precise World Control
Hugging Face researchers introduce H3-World, a framework that adapts the 33B MiniMax-H3 video generator into an interactive world...
Hi-Q: Evidence-Guided Query Refinement Boosts Multi-Hop QA Accuracy
Hugging Face researchers introduce Hi-Q, a framework that dynamically refines multi-hop queries into hierarchical trees based on c...