A new arXiv paper (2609.38406v1) introduces a ~9.2K-instance multi-domain benchmark for narrative text infilling and reports that model scale does not reliably predict infilling quality, with Gemma-2-…
A new paper from researchers at Google, the University of Chicago, and the University of London found that removing safety fine-tuning that tells AI models they aren't conscious makes them give more h…
A new training method called replacement-aware training produces sparse auto-encoders (SAEs) that retain language capabilities when used in a full replacement model of Gemma-2-2B, unlike standard SAEs…
Google hosted a Kaggle hackathon challenging developers to train non-reasoning Gemma-2-2B and Gemma-3-1B models into general reasoning models using Tunix and Kaggle TPUs. Over 11,000 entrants and 300+…