New research indicates that AI agents learn compressible models of data, which don’t have enough space to enable memorization.
How the Verus "program verifier", which automatically checks code against a mathematical specification of its functionality, helps increase security assurance in software projects.
Ten years after we founded the Automated Reasoning Group, mathematical logic has moved from academic research into production services that secure millions of customer workloads — demonstrating that systems can be provably correct, not just probably correct.
PatientAgentBench generates a synthetic patient health record, a realistic clinical vignette, and a patient agent that converses with the AI system under evaluation, to capture what a patient-facing agent actually has to do.
A new chiplet architecture, custom die-to-die connectivity, and support for DDR5-8800 memory and the latest PCIe gen6 interconnects improve performance by 25% for general-purpose and agentic AI workloads.
Customer-obsessed science
Research areas
-
September 25, 202612 min readUsing the Neuron Kernel Interface, a Reactor–AWS collaboration tackled the dynamic shapes, memory access patterns, and cache management that make real-time autoregressive diffusion hard—building techniques that generalize across models. -
September 21, 202611 min read -
-
August 21, 20269 min read
-
July 30, 20268 min read
Featured news
The Build on Trainium program is a $110 million credit initiative supporting AI research at 30 universities including Stanford, UC Berkeley, UIUC, UCLA, CMU, and MIT, with a focus on Responsible AI.
The Antibody Developability Benchmark is powered by one of the most diverse antibody datasets in public literature, enabling transparent performance evaluation for AI-guided antibody design.
Challenge pushes teams to demonstrate measurable gains in secure-coding performance while building AI agents that advance real-world utility and reliability at scale.
Initiative will fund over 100 doctoral students researching machine learning, computer vision, and natural-language processing at nine universities.
-
EMNLP 20262026Evolutionary prompt optimizers such as GEPA suffer from prompt bloat: each iteration appends rules and caveats, producing prompts up to 3×longer yet no more accurate. We trace this to three deficiencies - incomplete error observation, limited search diversity, and unreliable selection - and propose ESPO (Error-Structured Prompt Optimization), which decomposes prompt optimization into three phases: Diagnose
-
EMNLP 20262026Fine-tuning large language models (LLMs) for e-commerce attribute extraction requires labeled data representative across thousands of product types, attributes, and multiple languages. This combinatorial scale translates to millions of annotations, rendering human labeling prohibitively costly. While recent work has demonstrated synthetic label generation using LLMs (Negri et al., 2025), deploying such
-
2026Agent evaluation today depends on per-trace LLM-judge inference or human review, too expensive to run on every trace; production systems fall back to sampling a fraction. We find that failing agents leave a detectable behavioral signature in standard observability telemetry: disproportionate effort relative to outcome. We formalize four behavioral failure signatures from agent telemetry, validated on 4,671
-
2026Reinforcement-learning (RL) post-training of tool-using LLM agents leaves expensive rollout GPUs idle while each trajectory waits on CPU-side environment work: sandbox cold-start, code execution, retrieval, and external API calls. We term this idle the environment bubble and measure it directly on 8×A100 hardware using a 20 Hz NVML profiler that avoids the kernel-residency pitfall in the utilization counter
-
COLM 2026 Workshop on Lifelong Agents2026Multi-agent architectures today are hand-designed, generated one-shot, or evolved through unconstrained code search—none of which treats the number of agents as a searchable parameter, and none of which can discover that a single agent is optimal. We argue agent count and coordination topology should be optimized with the same discipline applied to neural architectures: bounded edits, held-out validation
Collaborations
View allWhether you're a faculty member or student, there are number of ways you can engage with Amazon.
View all
Gretchen Ertl
The program offers unrestricted funds and other resources to support research at academic institutions and non-profit organizations in areas that align with our mission.
A global university competition to drive secure innovation in generative AI technology, which focuses on responsible AI and large language model coding security.
Credit: Wolfram Scheible
We partner with particular academic organizations across the world for deep and sustained collaborations in multiple research areas of mutual interest.
Courtesy of Pai-Ling Yin
We hire world-class academics to work on large-scale technical challenges, while they continue to teach and conduct research at their universities.
