Latest stories
Hugging Face Paper Introduces Inverse Distillation Unlearning for One-Step Generative Models
A new framework called Inverse Distillation Unlearning (IDU) jointly distills multi-step diffusion and flow-matching teachers into...
New Post-Training Method Aligns Diffusion Language Models via Representation-Space MMD
Researchers propose a post-training objective that minimizes Maximum Mean Discrepancy between generated and reference distribution...
LoGRA Cuts LLM Reinforcement Learning Memory by Up to 45.7%
A new method called LoGRA compresses reinforcement learning signals into low-rank gradient sketches, cutting training memory by up...
OmniReasoning Targets the Blind Spot in Audio-Visual Joint Reasoning
A new Hugging Face paper introduces OmniReasoningBench, the OmniQA data engine, and a self-distillation method that together push...
Self-Generated Feedback Destabilizes Test-Time Training, New Study Finds
A new Hugging Face paper shows that when models learn from their own generated text during inference, prediction on independent hu...
Hugging Face Paper Proposes 3T Framework and Proactivity-Gym Testbed for Proactive LLM Agents
A new paper outlines three joint principles — Task Capability, Temporal Allocation, and Trust — for designing proactive LLM agents...
LMBuild Benchmark Tests Whether LLM Agents Can Build Objects That Actually Work
A new benchmark from Hugging Face researchers evaluates LLM agents on generating buildable, functional 3D structures rather than j...
MemAdapter Targets Memory-Induced Sycophancy in LLM Agents
A new framework called MemAdapter adaptively integrates retrieved memories so LLM agents can reason objectively instead of over-al...
ID-Forcing Tackles Drift in Long Autoregressive Video Generation
A new test-time framework, In-Distribution Forcing (ID-Forcing), keeps cached key-value entries in-distribution to stop color, tex...
Hugging Face Paper: LoHi Reframes Long-Video Efficiency Around Frames, Pixels, and Decoding
A new Hugging Face paper argues that long-video understanding should trade per-frame resolution for denser temporal coverage, and...
OSWorld-Pro Introduces Process-Based Evaluation for Computer-Use Agents
A new benchmark of over 300 tasks and 2,800 subgoals, grounded in 67,000 human annotations, evaluates computer-use agents on their...
CANOPY Brings Adaptive-Granularity Compression to Multimodal RAG
Hugging Face paper CANOPY proposes a post-retrieval compression framework that represents retrieved text, tables, images and video...