Top 10 Most-Viewed Publications of 2025

The top 10 most-viewed publications of 2025 show where Amazon Science's research attention was concentrated. The core pillars were the Amazon Nova family of models, AI safety, formal verification, and logistics and robotics.

The top rankings featured Amazon Nova Premier, Amazon Nova Sonic, and Amazon Nova 2 side by side. Nova Premier is a multimodal foundation model with a 1 million token context that handles text, image, and video, while Nova Sonic unifies speech and text into a single architecture aimed at low-latency voice interaction. Nova 2 is designed to balance speed and accuracy through dynamic reasoning and extended thinking controls, and consists of a family of 4 models including Omni and Sonic.

Safety and reliability also carried significant weight. Amazon's frontier model safety framework systematized procedures for identifying, evaluating, and mitigating risks for advanced frontier AI models, and Formally verified cloud-scale authorization rewrote AWS's authorization engine in Dafny to achieve mathematically verifiable correctness. The new engine was successfully deployed in 2024, with performance improving by 3x.

Real-world application cases stood out in logistics automation and user experience evaluation.

• Vulcan Pick: A robotic system that picks target objects from fabric pods, combining 3D scene understanding with adaptive motion control.
• Stow: A robotic packing system that places items into densely packed shelves, having processed over 500,000 stows at large-scale fulfillment centers.
• UXAgent: An LLM agent-based UX testing framework that pre-validates web designs using thousands of simulated users.

Also included were research extending statistical power calculations to effect size distributions for A/B testing, and research extending neural model checking to inductive invariants of safety properties. Overall, the popular papers of 2025 reveal how much interest was focused not just on generative AI itself, but on the technology for safely operating it and connecting it to real-world systems.