Researchers from Perplexity and NVIDIA have unveiled JOLT, a post-training technique in which a single model acts as both teacher and student. On the MATH benchmark it reached the same accuracy as GRPO while processing 6.5 times fewer training tokens.
NVIDIA has added new training tools, simulation checkpoints and Cosmos Data Factory blueprints to its Cosmos platform. World models help physical AI systems understand and simulate the real world, and the company says Cosmos ranks in the top five on six open model leaderboards.
This hands-on guide benchmarks Cactus Compute's 16.9 MB Whistle speech model against Whisper base on a CPU. It compares word error rate, speed, streaming latency, timestamps, keyword spotting and turning speech into tool calls.
Zurich-based startup ENACT has raised close to €27 million in pre-seed funding to build robots that better sense forces and adjust their movements when handling objects. Its approach is to design the mechanics, the data and the AI models for manipulation together.
A Hugging Face post covers ensemble forecasting with IBM's Granite time-series models, combining several models' predictions.
JDN ran 25 French queries three times each across four AI search environments and found large differences in the sources they cite. Anyone running a generative engine optimization strategy needs to know exactly what they are measuring, the article argues.