Build on Trainium: Agentic, recursive improvement of kernels and frameworks - Fall 2026
Build on Trainium: Agentic, recursive improvement of kernels and frameworks - Fall 2026 supports researchers developing self-improving AI agents to optimize Trainium workloads and reduce optimization time and cost.
Browse similar grants
Funder: Amazon Research Awards
Due Dates: October 1, 2026 (Submission opens) | November 4, 2026, 11:59 PM Pacific Time (Submission deadline) | February 2027 (Decision letters)
Funding Amounts: AWS Promotional Credits: up to $50,000, or up to $250,000 and beyond; one-time unrestricted awards
Summary: Supports self-improving AI agents that optimize kernels and end-to-end training and inference systems on AWS Trainium.
Key Information: Each PI may submit only one proposal per ARA call period; core research compute should primarily use Trainium.
Description
Amazon Research Awards’ Fall 2026 call supports agentic, self-improving frameworks for kernel and end-to-end workload optimization on AWS Trainium. It seeks systems that connect generation, evaluation, and learning: generating code, compiling and executing it, interpreting profiler feedback, checking correctness, and repairing or improving results. The call sits within Build on Trainium, a multi-year, $110 million credit initiative supporting AI research and university education.
Research priorities include kernel generation and auto-tuning, training and inference scheduling, memory planning, and optimization loops that reduce time and computational cost. Other areas include performance debugging and explainability, reward-hacking detection, correctness benchmarks, and supervised fine-tuning or reinforcement learning of open-source optimization models. Learning may improve model weights or agents’ context, tools, and knowledge bases, with an emphasis on reproducible work and community contributions.
