Amazon、Trainium 向け研究助成プログラム「Build on Trainium」の受賞者 34 件を発表

Build on Trainium is a $110 million credit program focused on AI research and university education aimed to support the next generation of innovation and development on AWS Trainium. The program provides compute credits to novel AI research on Trainium, investing in leading academic teams to build innovations in critical areas including new model architectures, ML libraries, optimizations, large-scale distributed systems, and more. This announcement includes awards funded under the Fall 2025 Build on Trainium: Responsible AI call for proposals. Proposals were reviewed for the quality of their scientific content and their potential to impact both the research community and society. This cycle’s focus on Responsible AI invited proposals addressing five priority topics: AI safety and alignment, multi-lingual language models, representation engineering, sustainability and small language models, and deep learning models for synthetic data generation—all leveraging AWS Trainium infrastructure. The recipients have access to more than 700 Amazon public datasets and can utilize AWS AI/ML services and tools through their AWS Promotional Credits, are assigned an Amazon research contact who offers consultation and advice, and benefit from AWS Trainium resources, such as tutorials and hands-on sessions. "Build on Trainium gives the next wave of AI researchers powerful, scalable access to Amazon's purpose-built AI chips, so the only limit is their imagination, not their compute budget," said Yida Wang, AWS AI Principal Applied Scientist. "By leveraging the support from Build on Trainium, University of Illinois Urbana-Champaign researchers are studying topology-aware parallelization strategies for large-scale mixture-of-experts models with as many as one trillion parameters on up to 1,024 Trainium chips. At the University of Washington, researchers are developing an inference-optimization framework that raises token efficiency for everyone building on Trainium, with the goal to deliver portable, high-performance LLM inference on Trainium." RecipientUniversityResearch titleWei BaoThe University of SydneyFACTOR: Federated Adversarial Co-Training with Textual Gradient for LLM Security and RobustnessViveck CadambeGeorgia Institute of TechnologyLeveraging Public-Private Mixtures For Differentially Private Synthetic Data GenerationYujun CaiThe University of QueenslandResponsible AI on Trainium: Scalable Detection and Mitigation of Evasive Multimodal Scam ContentHaipeng ChenCollege of William and MaryDELA: Editable Diffusion Language ModelsTianlong ChenUniversity of North Carolina at Chapel HillAlgorithm-System Co-Design for Efficient Sparse and Quantized LLMsSaadia GabrielUniversity of California Los AngelesMANSA: Democratizing Voice AI with Efficient Multimodal Foundation ModelsHaewon JeongUniversity of California Santa BarbaraLeveraging Public-Private Mixtures For Differentially Private Synthetic Data GenerationHaojian JinUniversity of California San DiegoGoverning Social Bias in AI Image Generation through Value ManifestsMarios KogiasImperial College LondonTowards Deterministic Model InferenceSachin KumarThe Ohio State UniversityNatively Multimodal and Multilingual Speech-Text Large Language ModelsEmanuele La MalfaInstitute for Decentralized AI (ADAI)Safe, Social Pre-training of LLM AgentsXiaoxiao LiThe University of British ColumbiaMemorization-Aware Preference Optimization for Machine UnlearningYingcong LiNew Jersey Institute of TechnologyEfficient and Adaptable Language Models via Sub-Model SearchZhijian LiuUniversity of California San DiegoAlgorithm-System Co-Design for Efficient Sparse and Quantized LLMsSongtao LuThe Chinese University of Hong KongM3-Align: Scalable Multilevel & Multiobjective Alignment for Multilingual Language ModelsYao LuUCL - University College LondonBreaking the Multilingual Data Wall: Scaling Synthetic Data for Low-Resource Language Model PretrainingSasa MisailovicUniversity of Illinois at Urbana-ChampaignCratos: Certified Robustness for Quantization and Pruning-Aware Training and Tuning of Vision Language ModelsTinoosh MohseninJohns Hopkins UniversityTRIM-LLM: From Quadratic to Linear Attention and Structured Pruning for Carbon and Cost-Efficient LLM Deployment on TrainiumThanhVu NguyenGeorge Mason UniversityLeveraging AWS Trainium for Verifiable AI and ML-Assisted Mathematical ReasoningFrank RudziczDalhousie UniversityRepresentation Immunization on Trainium: Scalable Noising & Weight-LockingAnuj SharmaIowa State UniversityBuild on Trainium: Physics-Grounded Synthetic Crash Generation for Vulnerable Road Users with Representation Engineering on Video Diffusion and VLMsShen ShenMassachusetts Institute of TechnologyAgent Tool-Use Safety Benchmarking with MCP-Specific LoRA MitigationsRyan ShiUniversity of PittsburghBenchmarking and Improving Multilingual LLMs on Real Indic Language Healthcare DialoguesNaichen ShiNorthwestern University LLM Hallucination Detection and Mitigation Jaideep Srivastava University of Minnesota Twin CitiesKnowledge-Infused Time-Series Pretraining with Safety-by-Knowledge-Checking for Trustworthy Clinical AICheng TanNortheastern University Towards Reliable and Trustworthy LLM Services with ϵ-correctnessYue WangUniversity of Central FloridaGame-Theoretic Frameworks for Responsible AI on Pluralistic AlignmentYang WangUniversity of Illinois at Urbana-ChampaignSafeguarding Youths in Multimodal Generative AI: Toward a Trainium-Powered Framework for Safety and AlignmentErmin WeiNorthwestern University Higher Order Based Fast LLM Training MethodJun WuMichigan State UniversityBigger Models, Bigger Risks? Investigating the Safety Landscape of LLM ScalingXiaokui XiaoNational University of SingaporeTrainium-Accelerated, LLM-Guided Differentially Private Synthesis of Hierarchical Relational DataMin XuCarnegie Mellon UniversityLanguage-Grounded Interpretability for ViT and 3D ModelsZiyu YaoGeorge Mason UniversityRepresentation Engineering of LLMs for Secure Code Generation Junzhe Zhang Syracuse UniversityDeconfounding Image Editing for Robust Causal Prediction