Hugging Face has released Transformers v5, marking five years since v4.0rc-1, with the ecosystem now featuring over 400 model architectures and more than 750,000 compatible model checkpoints, driven by widespread AI access and community contributions.

OpenAI and Snowflake have teamed up in a $200 million deal to integrate advanced AI capabilities into Snowflake’s enterprise data platform, allowing businesses to run AI agents and derive insights directly within their data environments.

Canvas in AI Mode is now available to all users in the U.S., providing a dynamic workspace for organizing projects and supporting creative writing and coding tasks.

DBench-Bio is a dynamic benchmark that evaluates large language models’ ability to discover new biological knowledge through a three-stage pipeline involving data acquisition, question-answer synthesis, and quality filtering. It addresses limitations of static benchmarks by ensuring evaluation content is unseen during training and remains up-to-date with rapid LLM developments.

DropMatch is a training-free method that uses Monte Carlo dropout on the LM head to evaluate draft tokens in speculative decoding, enabling adaptive acceptance based on semantic consistency without requiring training or data. It improves inference speed by up to 1.33x and integrates seamlessly with existing models and techniques.

GitHub rebuilt its search architecture for GitHub Enterprise Server to ensure high availability, improving reliability and performance for enterprise users.

The article explores deploying Vision–Language–Action (VLA) AI models on embedded robotic platforms like NXP i.MX95, emphasizing dataset consistency, asynchronous inference for real-time control, and hardware-aware optimizations to meet strict latency and power constraints.

A preprint demonstrates the extension of single-minus amplitudes to gravitons, using GPT-5.2 Pro to help derive and verify nonzero graviton tree amplitudes in quantum gravity.

Eedi, an online math tutoring service, uses AI to deliver personalized math lessons by assessing students through a dynamic diagnostic quiz and adapting questions based on responses to address specific learning gaps.

The paper proposes a ‘critic’ model trained from sparse, noisy real-world interaction data using rubric-based supervision, which improves coding agent performance in tasks like SWE-bench reranking and enables early stopping. It bridges the gap between academic benchmarks and real-world human-in-the-loop coding environments by leveraging observable behavioral features.