Amazon Science

Runtime optimizations and a thorough evaluation process driven by agentic AI help ensure that models written in PyTorch will execute efficiently and reliably on Trainium systems.

How Amazon partnered with AVERI to independently assess — and improve — our internal safety testing of Amazon models.

“Priming” reuses pretrained Transformer weights to build models with more efficient architectures at less than 1% of the original training cost and with throughputs up to twice as high.

Augmenting a network graph with agentic AI produces a “digital twin” that can help isolate network failures.

Using 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.

The collaboration aims to advance research while broadening participation and translating discovery into real-world solutions.

Three new papers from Amazon Bio Discovery address bottlenecks in AI-driven antibody engineering, from benchmarking binding predictors to experimentally validating de novo design.

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.

Discounting the opinions of LLM judges with highly correlated outputs ensures that panels of judges reflect a true diversity of perspectives.

Extendable framework enables testing agents on the full set of capabilities required to successfully complete a procedure, not isolated proxy tasks.

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.

A competition with a finalist ceremony during NeurIPS 2026, challenging researchers to train language models from scratch on Trainium, exploring what optimal architectures look like when the hardware changes.

Amazon announces 34 recipients of the Build on Trainium program, 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.

Instead of compromising among parameter updates dictated by different training objectives, ControlG allocates computational capacity to objectives sequentially and dynamically.

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.

As AI agents take on higher-stakes decisions, Lean programming language makes it possible to mathematically prove they will behave safely.

HydroShear, a new physics-based simulator, teaches robots how to use their sense of touch to perform complex manipulation tasks, in a way that transfers seamlessly to the real world.

A new Rust proxy called Turnstile sits between the model backend and the agent harness to capture information lost in mere text transcripts.

Amazon is developing precise, sector-specific approaches to measuring decarbonization progress — starting with emissions per unit shipped.