We present Gemini 2.0, a family of multimodal models achieving state-of-the-art performance across text, image, video, and audio understanding with native tool use and long-context… Read more →

This survey covers 340+ papers on world models applied to autonomous driving, categorizing approaches by architecture, training paradigm, and evaluation methodology, identifying ke… Read more →

We present Mamba-2, extending selective state space models with improved hardware-aware algorithms, achieving transformer-quality results at linear time complexity across language,… Read more →

DeepSeek-V3 demonstrates that open-weight models can match proprietary frontier systems through aggressive mixture-of-experts scaling, novel training stability techniques, and high… Read more →

We systematically quantify the capability loss incurred by RLHF and constitutional training across 12 frontier models, finding that modern techniques reduce the tax to under 2% on… Read more →

We demonstrate 2-bit quantization of 70B+ parameter models with less than 1% quality loss through a novel mixed-precision scheme, enabling frontier-class inference on consumer GPUs… Read more →

We introduce CAI 2.0, eliminating the need for human feedback in alignment training while achieving stronger safety guarantees through recursive constitutional evaluation and autom… Read more →

We revisit scaling laws for machine translation and find that previously established power-law relationships break down above 100B parameters, requiring new architectural innovatio… Read more →

We present the first comprehensive taxonomy of reward hacking failure modes in RLHF systems, documenting 47 distinct attack vectors and proposing a unified mitigation framework val… Read more →

We report on engineering and scientific lessons from training a 1T-parameter sparse MoE model, including novel load balancing, expert specialization dynamics, and inference serving… Read more →