MIT researchers unveiled PaTH Attention, a framework that replaces static positional encoding with content-aware transformations, enabling LLMs to track complex relationships over long texts with far less compute. Complementing this, MIT CSAIL’s DisCIPL system allows a “boss” model to orchestrate smaller, specialized models for intricate tasks like budgeting, while the SEAL framework boosts accuracy by 47% by teaching AI to generate its own “study notes.”
Nvidia accelerated the agentic AI race on December 15, 2025, with the Nemotron 3 family. These open models use a hybrid mixture-of-experts architecture to deliver 4x the throughput of previous versions. By launching the Nemotron Agentic Safety Dataset and NeMo Gym, Nvidia is providing the infrastructure for developers to build specialized, long-horizon agents for complex industrial use cases.
Regulatory tension peaked as Virginia legislators prepared a January push to restrict AI chatbots from mimicking human emotional support. This move, led by Delegate Michelle Maldonado, follows tragic reports of teen self-harm linked to AI interactions. Simultaneously, 42 state attorneys general have formally demanded that AI labs implement immediate safeguards to prevent chatbots from delivering sycophantic or dangerous responses to minors.
