Publications

We present Amazon Nova 2, a family of multimodal foundation models that demonstrate state-of-the-art performance in reasoning and generation tasks across text, image, and video modalities.

We present MIRAGE, a comprehensive benchmark for evaluating multimodal AI systems in the context of agricultural expert consultations, requiring complex information-seeking and reasoning capabilities.

We introduce Tree-of-Prompts, a framework for systematically optimizing prompts by abstracting the control-flow of prompt strategies into a searchable tree structure.

Technical report detailing Amazon Nova Premier, our most capable foundation model designed for complex reasoning, analysis, and agentic workflows.

We propose a real-to-sim-to-real pipeline that leverages vision-language models to generate iterative keypoint rewards for learning robotic manipulation skills.

We introduce the Amazon Nova family of foundation models, spanning multiple sizes and capabilities for diverse AI applications.

We present RoboEXP, a framework that constructs action-conditioned scene graphs through interactive exploration to enable more effective robotic manipulation.

We propose contextual data augmentation techniques that improve the robustness and generalization of task-oriented dialog systems.

We present methods for synthesizing code-switched text in language pairs not seen during training, enabling better multilingual NLP systems.

We present a comprehensive empirical analysis of cross-lingual phenomena in voice assistant interactions, using large-scale data from Alexa to understand how users naturally mix languages.