Model Genome: Fingerprinting Whether an LLM Was Trained From Scratch or Derived
mayafree
• • 16NEW Articles from Team or Enterprise organizations will get promoted to the main section.
GLInt: Geometry-Matched Hard Negatives for Late-Interaction Retrieval
chungimungi
• • 11Blackwell isn't a monolith: what happens when you actually measure kernel performance on GB10
curtburk
• From Point Clouds to Tensors: Bijective Gridification with SquareNet
adec314
• • 1On the integrity of the Arabic TTS Arena leaderboard
Navid-AI
• • 6Rebuilding Among AIs from Six Log Files
DedeProGames
• • 6Lattice: an 8 MB static retriever that embeds Wikipedia in 7 minutes
erikkaum
• • 19Decoder crowding predicts SAE steering side effects, and it is not just firing frequency
bishnoiyash
• Energy-Guided Flow Matching
xxlucas
• FP8 KV-Cache on Intel® Arc™ Pro B70: 2× Capacity with strong Long Context Throughput Gains
kushal-mittal
• • 4Publish, then improve: routing around the 20-Space cap
ParetoOptimal
• Reproducing 147 Machine Learning Papers in 16 Days on One Workstation
ParetoOptimal
• • 1Alpamayo 2 Super: The expert that didn't grow
JonnaMat
• Optimum Intel v2.1.0
echarlaix
• • 5IDEOGRAM-4 for inpainting with Modular Diffusers and Differential Diffusion
OzzyGT
• • 4Waves, queues, and the three free gates
ParetoOptimal
• Picking reproducible papers out of a 6,341-paper haystack
ParetoOptimal
• A $0 fine-tuned judge, and the eval that assigned its seat
ParetoOptimal
• Training a coding agent using the OpenCode harness in remote HF sandboxes with TRL and OpenEnv
sergiopaniego
• • 22Watching a pure-integer GPT think — one degree per frame
nmicic
•