Research Papers
Interactive Training 2: Open-Source Control Plane Enables Auditable Live Model Steering
Hugging Face introduces Interactive Training 2, an open-source control plane that allows humans and automated agents to adjust tra...
DataPrep-Bench: First Unified Benchmark for LLMs as Training Data Preparators
Hugging Face researchers introduce DataPrep-Bench, the first benchmark jointly evaluating LLMs' data construction and quality eval...
SceneActBench: New Benchmark Tests How Well VLMs Act on 3D Scenes
Hugging Face researchers introduce SceneActBench, a benchmark evaluating vision-language model agents on five 3D tasks within a un...
IDEAgent: Multi-Agent Framework Boosts Research Idea Quality and Diversity by 3.89x
Hugging Face researchers introduce IDEAgent, a multi-agent framework that treats research ideation as a Quality-Diversity search,...
Agentic Context Management: A New Framework for Agent Memory and Cost
Hugging Face researchers propose Agentic Context Management (ACM), a lifecycle-based approach to agent memory that treats context...
Hugging Face Study Reveals Scaling Laws for Native Multimodal Pre-Training
A new paper from Hugging Face investigates the scaling properties of native multimodal pre-training, showing that compute-optimal...
Molt: A Scalable PyTorch-Native Framework for Agentic RL
Molt is a new open-source, PyTorch-native training framework for agentic reinforcement learning that simplifies algorithm iteratio...
Skill Self-Play: Co-Evolving Skills Push LLM Capabilities to New Heights
A new framework called Skill Self-Play (Skill-SP) enables LLMs to co-evolve skills through a proposer, solver, and dynamic skill c...
DeepSeek's DSpark Boosts LLM Inference Speed by Up to 85% with Speculative Decoding
DeepSeek introduces DSpark, a speculative decoding system that accelerates LLM inference by 57–85% on V4, Qwen, and Gemma models w...
DeepSeek Unveils DSpark: Speculative Decoding Boosts LLM Inference by 60–85%
DeepSeek introduces DSpark, a speculative decoding framework that accelerates LLM inference by 60–85% while preserving byte-identi...
Kimi K2.5: Open-Source Multimodal Model with Agent Swarm and Vision-Coding Breakthroughs
Moonshot AI launches Kimi K2.5, a powerful open-source multimodal model with state-of-the-art coding and vision capabilities, feat...
Kimi Releases WorldVQA Benchmark to Test Visual World Knowledge in MLLMs
Moonshot AI's Kimi introduces WorldVQA, a benchmark evaluating factual visual knowledge in multimodal LLMs, revealing that even to...