So you want to see if your DP synthetic data method is actually any good. What makes a good benchmark? 1. Zero-shot performance should be low: the method should measure learning from the actual data;
2. Training on real data should work: learning should be possible; and… 2/n
MACHINE LEARNING
-
Benchmarking DP synthetic data: zero-shot and real data training
By
–
-
ContinuousBench: Hard Leakage-Proof DP Synthetic Text Benchmark
By
–
Does DP synth text transfer useful knowledge or just superficial style mimicking? Existing benchmarks: saturated Introducing ContinuousBench: a hard (curr methods fail at ε=100! ) & leakage-proof benchmark for DP synth text! Followup to our #ICML2024 best paper 1/n
-
Benchmarking confirms reasoning-only model requires thinking mode
By
–
Awesome, thanks for benchmarking it! It makes sense, the model is reasoning-only so thinking mode is mandatory here.
-
Synthetic Data as Key to Making Robots Truly Usable
By
–
synthetic data is the key to make robots really usable
-
New doc page: rendering Agent Traces on Hugging Face Hub
By
–
new doc page dropped! rendering Agent Traces on the @huggingface Hub
-
NVIDIA Launches Cosmos Coalition for Open World Models
By
–
7/ NVIDIA also launched the Cosmos Coalition – with Black Forest Labs, Runway, Skild AI, Agile Robots, LTX and Generalist – to push open world models forward together. The bet: world models are becoming the intelligence layer for robots and AVs, and NVIDIA wants the open
-
Cosmos 3 Nano and Super models on Hugging Face with datasets
By
–
6/ Two sizes, both live on Hugging Face right now: Cosmos 3 Nano (8B) — runs on a single workstation GPU for real-time robotics
Cosmos 3 Super (32B) — datacenter-grade, max quality Plus six open datasets and full post-training scripts on GitHub. -
Cosmos 3 tops open leaderboards for text, video, physics, and robotics
By
–
5/ And it's not a demo. Cosmos 3 tops the open leaderboards: #1 open model on Artificial Analysis for text→image AND image→video
#1 on Physics-IQ for physics accuracy — ahead of Sora 2
Leads PAI-Bench overall, ahead of Veo 3.1
#1 robot policy on RoboArena
Open weights beating -
Cosmos 3: Synthetic data generation for robotics, 60x faster, including rare edge cases.
By
–
4/ The real unlock is data. Robotics has always been bottlenecked by how little real-world training data exists. Cosmos 3 generates physically-accurate synthetic data up to 60x faster – including the rare edge cases you can't safely film: collisions, near-misses, accidents.
Eval -
Cosmos 3: native action generation for VLM, world model, robot
By
–
3/ Inputs and outputs span text, image, video, audio AND action.
— Chubby♨️ (@kimmonismus) 1 juin 2026
That last one is the big deal. Cosmos 3 was trained natively to generate actions, so the same checkpoint can run as a vision-language model, a video world model, or a robot policy. No multi-model orchestration. pic.twitter.com/PoLsK33ytZ3/ Inputs and outputs span text, image, video, audio AND action. That last one is the big deal. Cosmos 3 was trained natively to generate actions, so the same checkpoint can run as a vision-language model, a video world model, or a robot policy. No multi-model orchestration.