§ BEAT
Research
Microsoft's OpenForgeRL Trains Agents in Production Harnesses
Microsoft shrinks pathology AI model by 50×, enabling hospital deployment
Dense Patch Tokens Match Vision-Language Models at 1% the Parameters
Three Hours on $329 GPU Replaces Thousands of Hours of NAS Training
Training-Efficient Low-Rank Compression Sidesteps Serving-Speed Proof
Hugging Face Cuts Inference Attention Overhead 20-40% With Fused Kernels
DynaKRAG Boosts Multi-Hop QA Accuracy by Up to 5.78 Points
PAW Trades Compile Time for 1/50th the Inference Memory
ReContext Fixes Long-Context Retrieval Without Retraining Models
Researchers Close Gap Between AI Agents and Hand-Curated Skills
AI Agents Double Repository-Level Merge Friction
Open-Weight Pipeline Achieves 68% Accuracy Extracting Political Networks from News
OpenThoughts-Agent Dataset Hits 44.8% on Agentic Benchmarks
Moebius Model Reaches Browser via ONNX+WebGPU in Parallel Agent Session
Princeton Releases LOCUS, Machine-Readable Corpus of 9,239 US Local Ordinances
Single Dense Model Hosts Hundreds of Agent Personas as Lightweight Masks
Component Interaction, Not Quality, Determines Agent Performance
Agents-K1 Replaces RAG Text Chunks With Typed Scientific Knowledge Graphs
Tahoe Text-to-SQL System Cuts Compiler Feedback by 96%
EEVEE Surpasses Self-Improving Agents with 48% Margin on Multi-Domain Inference
Piper Compiler Eliminates Hand-Coding for Distributed Training
FASE Cuts Hallucination Detection to 333x Speed
SIGA Speeds Coding Agents on Scientific Simulators by 36×
Output Format Drives Faster Accuracy Loss Than Domain Shift in Multimodal LLMs
GPIC Open-Source Dataset Displaces ImageNet-1K as Standard Training Corpus
Omega-QVLA Cuts Robot Vision Model Memory by 71% Without Retraining
Production Hardware Tests Needed Before OFT Replaces LoRA at Scale
Schema.org Metadata Cuts Agentic Retrieval Errors by Two-Thirds
Meta Shrinks Mixture-of-Experts to Smartphones Without Cloud Offloading
