Source archive
Amazon Science publishes research from teams working across machine learning, natural language processing, computer vision, robotics, and operations. Its AI work is connected to real-world applications in commerce, logistics, cloud services, and consumer products.
Amazon is investing in the Lean Focused Research Organization, which leverages the Lean programming language to mathematically prove the safety of AI agent behavior. This approach is increasingly important as AI agents are used in higher-stakes decision-making.
Why it matters: The investment supports efforts to mathematically verify AI safety, addressing crucial risks as AI agents are deployed in sensitive environments.
Amazon and the University of Michigan have developed HydroShear, a physics-based simulator that teaches robots to use tactile sensing for complex manipulation tasks. The approach is designed to transfer seamlessly to real-world applications.
Why it matters: This advancement could enable robots to perform delicate tasks requiring a sense of touch, expanding their utility in manufacturing and other industries.
Amazon Science has developed Turnstile, a Rust proxy that sits between the model backend and the agent harness to capture information that is lost in plain text transcripts during agentic interactions. This enables the preservation of token IDs, which can support improved reinforcement learning.
Why it matters: Capturing token IDs directly provides richer data for reinforcement learning in agentic systems, potentially enhancing model training.
Amazon Science announced Graviton5, a new processor featuring a chiplet-based architecture, custom die-to-die connectivity, DDR5-8800 memory, and PCIe Gen6 interconnects. The chip delivers a 25% performance improvement for general-purpose and agentic AI workloads, while also increasing energy efficiency beyond Moore's Law.
Why it matters: Graviton5 marks a significant advancement in custom silicon for cloud computing, enhancing AI inference and general-purpose workloads in AWS data centers.
Amazon Science reports that splitting the 'separation kernel' from the rest of the Nitro security system and coding it in a subset of Rust enabled formal verification. This provides mathematical assurance of virtual-machine isolation on EC2.
Why it matters: Formal verification of the isolation engine provides mathematical assurance of VM separation, a critical security property for cloud computing.
Amazon Science discusses the challenges of automatically fact-checking long, AI-generated research reports, emphasizing that ground truth should be seen as an ongoing process rather than a static dataset. The article also notes the need for new benchmarks to address these challenges.
Why it matters: This perspective encourages continuous verification in evaluating AI-generated content, rather than relying solely on fixed datasets.
Amazon Science reports that AWS is exploring the use of flat network topologies in its data centers, utilizing quasi-random designs and new passive optical components called ShuffleBoxes. These innovations aim to make flat networks as practical and efficient as traditional fat-tree networks.
Why it matters: This development could improve data center network efficiency and potentially reduce costs for cloud computing infrastructure.
Amazon has announced the recipients of its Fall 2025 Research Awards, representing more than 49 universities across 11 countries. Awardees will receive access to Amazon public datasets and AWS AI/ML services and tools.
Why it matters: This program supports academic research in AI and machine learning, fostering collaboration between industry and academia.
Amazon Science researchers have introduced a new scaling law that connects specific architectural choices in large language models (LLMs) to their loss, allowing for the identification of models that can improve throughput by up to 47% without any loss of accuracy. This approach enables more efficient LLM inference while maintaining performance.
Why it matters: This scaling law provides a systematic method to accelerate LLM inference, potentially reducing costs and latency in production systems without sacrificing accuracy.
Amazon Science introduces Promptimus, an automated framework that refines existing LLM prompts by targeting specific failure points. The system enhances prompt performance without manual engineering or compromising existing functionality.
Why it matters: This reduces the need for manual prompt tuning, making LLM deployment more efficient and accessible.
Amazon engineers and scientists have developed new tools to optimize delivery networks under uncertainty, allowing for continuous adaptation. These tools specifically target the middle-mile network, which handles the movement of packages between fulfillment centers and delivery stations.
Why it matters: This development could enhance the efficiency and reliability of Amazon's logistics operations.
Amazon Science describes how its researchers reproduced three attacks capable of extracting private training data from AI models, as well as the cryptographic defenses that can prevent such breaches. The work underscores ongoing efforts to enhance data privacy during AI training.
Why it matters: This research highlights practical approaches to defending against data extraction attacks, which is crucial for maintaining privacy in AI systems.
Amazon Science researchers have introduced a statistical framework to estimate the likelihood of catastrophic failures in large language models during adversarial conversations. This method enables quantification of risks associated with LLM interactions.
Why it matters: The framework provides a systematic way to assess safety risks in LLMs, which is important for their deployment in sensitive contexts.
Amazon Science highlights Isabelle/HOL as the proof assistant that enabled the world's first formally verified cloud hypervisor, the Nitro Isolation Engine. The tool's balance of expressiveness, automation, and scalability was key to this achievement.
Why it matters: This marks a significant milestone in cloud security, demonstrating that formal verification can be applied to critical infrastructure at scale.
Amazon Science reports that a single, optimized large language model (Amazon Nova) now unifies molecular-property prediction tasks that previously required multiple models. The model can serve as a reasoning partner for medical chemists in drug discovery.
Why it matters: This advancement could accelerate drug discovery by providing a unified AI tool for molecular property prediction, reducing the need for multiple specialized models.
AWS and the Gray Lab at Johns Hopkins Whiting School of Engineering have announced the Antibody Developability Benchmark, a database for AI/ML antibody design. The benchmark is powered by one of the most diverse antibody datasets in public literature, enabling transparent performance evaluation for AI-guided antibody design.
Why it matters: This benchmark provides a standardized, transparent dataset for evaluating AI models in antibody design, potentially accelerating drug discovery and development.
Amazon's RuleForge system uses agentic AI to generate production-ready detection rules 336% faster than traditional methods. This system operates at a global scale, improving vulnerability detection across Amazon's infrastructure.
Why it matters: This highlights a practical application of agentic AI in cybersecurity, accelerating threat detection and response at scale.
Amazon Science describes how automated reasoning is used to balance security, performance, and maintainability in post-quantum cryptography. This approach helps ensure that cryptographic implementations are both correct and efficient.
Why it matters: As quantum computing advances, verifying and optimizing post-quantum cryptography is important for maintaining secure cloud services.
Amazon Science describes the use of techniques such as low-rank adaptation, data augmentation, and chain-of-thought reasoning to enhance LLM-based text-to-speech systems. These approaches support accent-free polyglot outputs, greater expressiveness, and more reliable speech synthesis.
Why it matters: This research could make AI-generated speech more natural and adaptable for diverse global applications.
Amazon Science describes how mechanism design theory, specifically an approach called agentic mechanism, enables Amazon and its vendors to optimize supply chain management without disclosing private information. This method allows for efficient collaboration while maintaining data privacy between parties.
Why it matters: This highlights a real-world application of economic theory to improve supply chain efficiency while protecting sensitive data.