Unified Latents Hits 1.4 FID by Replacing Stable Diffusion's Ad Hoc VAE with a Diffusion Prior
Infrastructure
23 posts in Infrastructure
80.3 on ScreenSpotPro: GUI-Owl-1.5 Sets New Bar for Open-Source GUI Agents
3.5x Faster Image Generation: DDiT Dynamically Resizes Patches in Diffusion Transformers
Reasoning Overthinking Solved: SAGE Cuts Tokens 44% While Improving Accuracy
Baidu Introduces ERNIE 5.0: Trillion-Parameter Unified Multimodal MoE Rivals GPT-5
First Holistic OCR Model: OCRVerse Unifies Document Parsing and Code Generation
175% Faster Prefill with Better Accuracy: ConceptMoE's Adaptive Token Compression for MoE
6x Fewer Tokens, Better OCR: DeepSeek's Visual Causal Flow Beats GPT-4o and Gemini
2x Faster VLA Inference with 70% Fewer Layers: Shallow-π Distillation for Edge Robotics
UPLiFT vs Cross-Attention Upsamplers: Linear Scaling Meets SOTA Quality
90% Attention Sparsity with Zero Quality Loss: SALAD Speeds Up Video Diffusion 1.7x
FP8 Rollout Instability Solved: Jet-RL Unifies Precision for Stable RL Training
73% on BrowseComp: Meituan's 560B Open-Source Model Leads Agentic Benchmarks
Tsinghua Researchers Show Diffusion LLMs Reason Better When You Take Away Their Flexibility
97ms First-Packet Latency: Qwen3-TTS Beats ElevenLabs in Voice Cloning Across 10 Languages
56.7% on OSWorld: EvoCUA's Evolutionary Training Beats Closed-Source Computer Use Agents
Microsoft's Agent Lightning Decouples RL Training from Agent Logic, Enabling Fine-Tuning of Any AI Agent with Zero Code Changes
16x Faster On-Device Video Generation: Qualcomm's ReHyAt Distills Attention in 160 GPU Hours
Gold Medal at IMO and IOI: DeepSeek-V3.2 Matches GPT-5 with Open Weights
6% Better Math Reasoning in Fewer Tokens: Multiplex Thinking Merges Multiple Paths into One
Zero Training Data, Full Performance: Dr. Zero Matches Supervised Search Agents
Why Reasoning Models Cheat on Efficiency: TNT's Fix Cuts Tokens 50%
Why Hyper-Connections Explode at Scale: DeepSeek's Manifold Fix
