Amazon has recently announced the latest cohort of Amazon Research Awards (ARA) recipients. This year's list includes 63 awardees from 41 universities and research institutions across eight countries, with at least 26 Chinese scholars making the cut.

Their research interests are diverse, covering AI safety, advertising systems, agentic AI, the in-house AI chip Trainium ecosystem, and the ambitious frontier science exploration dubbed 'Think Big.'

If the previous phase of AI development was centered around model parameter competition, then this year's list can be seen as Amazon's 'technology radar map' for the next stage of the AI battlefield. Safety, agentic AI, computing infrastructure, and interdisciplinary studies are all prominently featured in the awarded projects.

Established in 2015, the Amazon Research Awards program essentially serves as Amazon's long-term 'seed funding' initiative for academia. Recipients receive one-time research grants, AWS computing resources (including AWS Promotional Credits worth tens of thousands of dollars), access to over 700 Amazon public datasets, and internal Amazon research contacts to foster collaboration and facilitate project implementation.

This round (Spring 2025) encompasses five key solicitation directions:

• Amazon Science
• AI for Information Security
• AWS AI: Agentic AI
• Build on Trainium
• Think Big

In this technological landscape, Chinese scholars are represented at nearly every critical node.

AI for Information Security Direction: Eight researchers were selected, including three Chinese scholars:

• Zhou Li (University of California, Irvine)
  Zhou Li focuses on leveraging large models for 'readable and traceable' attack attribution in complex audit logs. Beyond merely detecting anomalies, his work ensures that security analysts can comprehend and trace back the findings.
• Yu Meng (University of Virginia)
  Yu Meng's project, titled 'Weakly Supervised RLHF,' aims to explicitly model ambiguity and uncertainty in capturing human preferences. Most alignment technologies currently assume that human preferences are clearly definable. However, in reality, annotator judgments are often influenced or even self-contradictory. Encoding this 'ambiguity' into the training process is an attempt to prevent large models from 'overfitting' at the value level.
• Ziming Zhao (Northeastern University)
  Ziming Zhao focuses on 'how large models attack and how to detect and repair these vulnerabilities in an explainable manner.' As large language models (LLMs) are integrated into more security-sensitive systems (ranging from code generation to automated penetration testing), the models themselves can serve as both tools for defenders and new attack surfaces. Zhao's team aims to clearly dissect 'how LLMs can be used for hacking' and then construct defensive measures accordingly.

Among the awardees in this round, Agentic AI emerged as one of the directions with the most recipients, with 30 researchers funded in this category alone, including a significant proportion of Chinese scholars.

• Cong Chen, Dartmouth College
  Starting from power systems, Cong Chen works on 'decision-making grid agents.' In grids with increasing renewable energy penetration, his research enables agents to perform scheduling and pricing for operators in complex market environments, balancing safety, cost, and carbon emissions.
• Chunyang Chen (Technical University of Munich)
  Chunyang Chen's research lies at the intersection of software engineering, human-computer interaction, software security, and AI. His main research directions include AI and large language model-assisted automated mobile application development, AI-enabled software repository mining, and deep learning and mobile application security.
• Bang Liu, Université de Montréal & Mila
  Bang Liu's project focuses on 'foundational protocols for collaborative agents,' while applying agents in materials science and multimodal learning. He previously advanced the integration of large materials science models with agents for designing new materials. The awarded project further explores what 'languages' and 'rules' should be used when multiple agents collaborate.
• Lianhui Qin, University of California, San Diego
  Lianhui Qin researches ReaL-Agent—a cross-modal agent that deeply integrates retrieval and reasoning. In multimodal scenarios, it must not only retrieve the correct graphic text (image-text) information but also follow a reasonable reasoning path step by step, avoiding 'retrieval without thinking.'
• Jindong Wang, College of William & Mary
  Jindong Wang focuses on a seemingly abstract yet critically important question: How important is the task 'structure' for LLM agents? Simply put, it's about designing more suitable task topologies for agents to enable more efficient decomposition, planning, and collaboration in complex tasks. This aligns with his long-term work in transfer learning, foundational models, and generative AI—advancing from 'can models learn?' to 'can models be effectively used in complex social scenarios?'

The Think Big direction has only three awardees, all deemed to possess 'transformative potential' in research vision, including one Chinese scholar:

• Tianlong Chen, University of North Carolina at Chapel Hill
  Tianlong Chen's project attempts to combine molecular dynamics simulations with protein AI models. In layman's terms, it enables AI models to not only observe the static structure of proteins but also understand their motion and interactions over time, thereby more accurately predicting folding, binding, and function. Such research both continues along the trajectory of works like AlphaFold and echoes AWS's internal emphasis on healthcare scenarios—from drug screening to disease mechanism modeling, high-precision protein models serve as infrastructure-level assets.

References:

https://www.amazon.science/research-areas/latest-news/63-amazon-research-award-recipients-announced-spring-2025

https://mp.weixin.qq.com/s/vq8r54VNGiiV67ykYyQfqA