마이크로소프트연구소와 공동연구협력을 위한 프로젝트 모집 공고

1. 목적

마이크로소프트연구소와 공동연구협력 프로젝트* 선정

* 2026년도 「디지털분야글로벌연구지원사업(과학기술정보통신부)」중 ‘기업(연구소)연계형’ 과제로 추진되며, 본 프로젝트는 ‘국가연구개발혁신법 제9조(예고 및 공모 등) 제4항, 동법 시행령 제9조(연구개발과제 및 연구개발기관의 공모 절차)’에 해당함.

2. 운영방향

마이크로소프트연구소에서 선정한 연구주제에 부합하는 창의적 아이디어 공모를 통해 선정

프로젝트 수행시 공동연구협력을 위한 마이크로소프트연구소의 전문가 매칭

공동 연구원* 마이크로소프트연구소에 180일 파견 (파견시점 파견 국가 규정에 따라 변경 가능)

* 공동연구기관(국내 대학)의 연구원(석박사생)

3. 과정개요

A. 프로젝트 지원

– 지원규모: 총 30억 원*, 25-30개 과제

* 디지털분야글로벌연구지원사업(과학기술정보통신부) 예산 : 30억 원

– 지원분야*:

* Scalable AI Systems and Infrastructure Optimization (‘26년 신규 과제): 14번~24번까지 해당, 9-11개 과제 선정 예정

1. AI × HCI: Toward Human–Agent Collaboration

As AI systems increasingly act as autonomous collaborators rather than passive tools, understanding how humans and agents can jointly perceive, reason, and create becomes a central challenge. By integrating perspectives from AI, HCI, cognitive science, and social interaction design, we aim to develop principles and systems that enable effective, trustworthy, and transparent human–agent collaboration. Our goal is to advance the science and design of interactive intelligence—where humans and agents co-adapt, share goals, and achieve outcomes neither could realize alone.

Topics:

Human–agent collaboration and co-adaptation
Shared autonomy, control, and goal alignment
Trust, transparency, and explainability in agent interaction
AI-augmented human cognition, creativity, and problem-solving
Design methods for mixed-initiative and co-creative systems

2. Media-Centric Intelligence & Agentic Memory

In the era of AI, media, such as images, video, audio, and text, has become the primary source from which intelligent systems acquire knowledge and generate insight. Media AI focuses on enabling AI systems to learn from and interact with real-world media sources. Achieving this requires transforming the inherent complexity and noise of the real world into abstract representations that capture essential semantics and underlying dynamics. Neural compression plays a key role in efficiently processing long-range contexts and sparking the development of advanced intelligence. We also focus on multimodal understanding for deep comprehension and multimodal generation for producing rich, context-aware outputs. This synergy ensures seamless integration and interaction across heterogeneous data types, significantly boosting the system’s intelligence and adaptability. The resulting AI system natively interacts with media, delivering efficiency, transparency, and practical deployment, ultimately driving progress in large-scale multimodal AI models for greater capability and reliability.

Topics:

Multimodal representation learning
Neural compression
Agentic memory and knowledge learning
Multimodal understanding
Multimodal generation and word model
Interactive UI generation

3. AI-Transformed Medical Education and Research

As artificial intelligence continues to reshape the healthcare landscape, the medical field stands at the forefront of this transformation. The integration of AI into healthcare presents both unprecedented opportunities and complex challenges that demand close interdisciplinary collaboration. Our mission is to develop solutions that prepare future medical professionals for evolving healthcare needs, drive cutting-edge research, and accelerate the translation of innovation into real-world clinical practice. We invite professors and researchers from diverse domains—including computer science, medicine, education, and ethics—to join us in this critical endeavor. Together, we can shape the future of AI-powered medical care and education, ensuring that technological progress translates into meaningful benefits for patients and society.

Topics:

Medical Foundation Models
Medical Agentic System
AI-Transformed Medical Education

4. AI and the Brain

We invite collaborators to join us in forging a synergistic relationship between Artificial Intelligence and the human brain. To unlock the next frontier of technology and neuroscience, we believe that integration of expertise in AI and neuroscience is essential. By bridging this critical gap, we can uncover transformative opportunities that will shape the future of human–machine interaction and cognitive health, such as designing AI systems modeled on the brain’s architecture for adaptability and efficiency, developing direct communication pathways between neural activity and external devices, and leveraging AI to advance brain research, diagnostics, and treatments for neurological conditions. These domains are not only groundbreaking in science and technology but are of profound importance for society and humanity. Together, we can redefine what it means to merge human intelligence with artificial intelligence for better lives, healthcare, and innovation.

Topics:

Brain-Inspired AI
Brain-Computer Interfaces
AI for Neuroscience / Brain Health

5. Societal AI

In an era marked by rapid advancements in AI, the impact of these technologies on society encompasses a broad spectrum of challenges and opportunities. To navigate this complex landscape, it is vital to foster an environment of interdisciplinary research. By integrating insights from computer science, social sciences, ethics, law, and other fields, we can develop a more holistic understanding of AI’s societal implications. Our goal is to conduct comprehensive, cross-disciplinary research that ensures the future trajectory of AI development is not only technologically advanced but also ethically sound, legally compliant, and beneficial to society at large. (more information can be found here)

Topics:

AI’s impact on human cognition, learning, and creativity
Ensuring AI safety, reliability, and control
Aligning AI with human values and ethics
AI’s transformation of social science research
Evaluating AI in unforeseen tasks and environments

6. Self-Evolving Agents

In the era of experience, agents should improve through sustained interaction and deployment feedback. Different from conventional learning techniques which require large scale human-crafted data, intelligent agents should be able to continually learning from its own experience generated via interactions with the world, just like our human beings. The framework Agent-lightning (https://github.com/microsoft/agent-lightning) developed by our group, designed for agent optimization, provides a good foundation here. Toward the goal of enabling self-evolving agents, innovations from both algorithm and system side are required. Algorithmically, some fundamental aspects, including advancing long-horizon decision making with principled credit assignment, reliable off-policy training, and safe learning from heterogeneous experience, multi-agent learning, play critical role in enabling efficient and effective evolving agents in real-world applications. Systemically, innovating on Agent-lightning for high-throughput experience generation, continual training, and extensibility are critical for large-scale stable optimization. The scope is intentionally broad — from digital agents (search, coding, dialog, enterprise workflows) to physical embodiments (vision-language-action robots and other embodied systems) — so methods and metrics transfer across modalities.

Topics:

Rewards and credit assignment for long-horizon tasks and its theoretical understandings
Multi-agent learning framework
Memory in agent learning
Learning from logs, interventions, and online feedback with safety considerations
System co-design on Agent Lightning: data pipelines and tracing, efficiency,
Cross-domain applicability: from digital agents to physical intelligence

7. Sample Efficient Embodied AI with Reinforcement Learning

Embodied AI systems often require extensive data to learn effective behaviors and show limited generalization in real-world deployment. We propose Sample Efficient Embodied AI with Reinforcement Learning as a collaboration theme focused on developing agents that learn with minimal data. By integrating model-based RL, world models, action abstraction, and efficient exploration, we aim to create embodied agents that acquire skills quickly and generalize across tasks with far fewer interactions. This direction targets practical embodied learning and high-impact scientific contributions.

Topics:

World models and model-based RL for reducing interaction cost
Action abstraction and hierarchical control for faster learning
Efficient exploration and representation learning in embodied domains
Sample-efficient transfer to novel tasks, objects, and environments
Benchmarks and evaluation protocols for data-efficient embodied AI

8. Spatial Intelligence

We live in a three-dimensional physical world, and Spatial Intelligence stands as a critical frontier in the evolution of AI, demanding not only a deep understanding of three-dimensional environments but also the capacity to act effectively within them. Unlike existing foundation models that focus on static data in language or 2D vision, spatially intelligent systems require reasoning in 3D — analyzing the rich spatial information of the environment, imagining how an environment changes, and choosing actions to accomplish meaningful goals. Our vision is to develop foundation models that unify perception, generation, reasoning, and action, enabling digital agents and real-world robotics to generalize across diverse environments and tasks with minimal adaptation.

Topics:

Spatial Large Language Model (LLM) and Vision-Language-Model (VLM)
3D Vision Foundation Models for Reconstruction and Generation
Robotic Vison-Language-Action (VLA) Models
Dexterous Hand Manipulation
Data and Benchmark for Spatial AI
3D World Models
Reinforcement Learning for Manipulation

9. Embodied Robotics Foundation Models

As robotics advances from controlled laboratory environments toward real-world deployment, Embodied AI has become a cornerstone of next-generation intelligent systems. Embodied Robotics Foundation Models integrate multimodal perception, 3D spatial understanding, action generation, and robot learning at scale—enabling robots to perform diverse tasks across different embodiments and environments.

To unlock their full potential, interdisciplinary research is essential. Progress in robotics hardware, control theory, computer vision, large-scale learning, cognitive science, and safety engineering must converge to build intelligent systems that are not only capable, but also robust, safe, and trustworthy. Our goal is to develop a unified foundation for robots that can reason, act, and adapt in the physical world, ultimately enabling general-purpose robotic assistants that benefit industry and society.

Topics:

Large-scale Vision-Language-Action (VLA) and world models for robotics
Cross-embodiment and cross-scene generalization (robot morphology, environment, tasks)
Learning universal latent action spaces and transfer to real robot actions
Few-shot and zero-shot task learning, adaptation, and long-horizon planning
Safety, controllability, and reliability in embodied action generation
Integration of multimodal sensing (3D perception, tactile, proprioception)
Data collection at scale: teleoperation, simulation, and synthetic data
Evaluation of robots in unconstrained real-world environments

10. Multimodal Agents for Digital Embodiment

In the current digital era, the sheer volume of information and the complexity of digital interfaces have become a significant bottleneck for human productivity and research efficiency. We are moving beyond passive computer vision to a paradigm of Grounded Digital Autonomy. This research area focuses on building autonomous agents that possess sophisticated visual intelligence to perceive, reason over, and act within dynamic digital environments, including operating systems, web browsers, and complex visual documents. By integrating deep reinforcement learning, advanced visual grounding techniques, multi-modal reasoning, and persistent memory systems, our goal is to enable machines to become truly intelligent research and task assistants, capable of automating long-horizon goals, synthesizing new knowledge, and providing high-quality critique of dense visual data. This research is critical for redefining human-computer interaction and achieving a new level of informational efficiency.

Topics:

Foundational VLM Architectures for Robust and General Visual Reasoning
VLM-Powered Semantic Analysis, Critique of Visual Documents, and Deep Visual Grounding of Complex Graphical User Interfaces (GUIs)
Agentic Planning and Reasoning for Long-Horizon Information Synthesis
Reinforcement Learning frameworks for Tool Orchestration and Complex Digital Task Completion
Persistent Agentic Memory and Knowledge Retention Architectures
Hybrid Agent Architectures integrating VLM, Code Execution, and System APIs
Self-Improvement Mechanisms for Autonomous Agents through Performance Feedback and Reflective Adaptation
Meta-Learning Paradigms for Agent Adaptability and Cross-Task Skill Transfer

11. Multimodal Agent

As AI evolves from perception to action, the next frontier is creating agents that can reason across modalities and operate effectively in both physical and digital environments. Multimodal agents integrate vision, language, and action with predictive world models to achieve robust, adaptive behavior in dynamic contexts. By combining advances in machine learning, robotics, and reasoning algorithms, we aim to develop systems that understand, plan, and execute complex tasks safely and efficiently. Our goal is to establish foundational research that enables generalist agents—unlocking applications in automation, assistive robotics, and digital services.

Topics:

Multimodal Reasoning for Grounded Decision-Making
Vision-Language-Action Models for Instruction-Following and Dexterous Skills
Efficient Vision-Language Models
World Models for Predictive Planning and Sample-Efficient Learning

12. Spatial Intelligence

Topics:

Spatial Large Language Model (LLM) and Vision-Language-Model (VLM)
3D Vision Foundation Models for Reconstruction and Generation
Robotic Vison-Language-Action (VLA) Models
Dexterous Hand Manipulation
Data and Benchmark for Spatial AI
3D World Models
Reinforcement Learning for Manipulation

13. Scientific AI

AI is transforming the way scientific knowledge is generated, analyzed, and applied. By integrating AI into research workflows, we can enhance understanding, improve predictions, and accelerate innovation across disciplines. This theme explores AI as a powerful partner in scientific discovery, capable of augmenting human reasoning, guiding experiments, and enabling new forms of exploration.

Topics:

AI for clean energy and sustainability
AI-powered physical simulations
AI Scientists: AI systems that autonomously generate hypotheses, design experiments, and support scientific reasoning
Human–AI collaboration and explainable scientific reasoning

14. Agentic AI for System Research

Current AI can now write code—but not systems. “System Intelligence” has been used to refer to the fundamental leap on AI’s capabilities to design, implement, and operate next-generation computing systems. Realizing system intelligence must unlock AI’s next stage of capabilities – reasoning architectures and protocols, weighing tradeoffs, applying enduring principles, developing effective abstractions, and more – beyond coding and bug fixing.

We are actively exploring this direction and studying key technologies, such as future agentic systems and system reasoning models. We believe that realizing system intelligence requires the collective efforts of the entire system research community and must be accomplished atop the decades of systems research expertise and experience (much of which today’s AI has never seen in its training data).

Topics:

Desigen future agentic systems
Improve model’s system reasoning capabilities
Reinforcement learning with system feedback

15. System Support for Large Context RL Post Training of Current & Future LLMs

Reinforcement learning (RL) post training is emerging as a critical pathway for improving LLM reasoning—especially for multi turn dialogue, chain of thought refinement, tool use, and decision making under uncertainty. However, RL setups fundamentally stress the context budget: trajectories span many turns; verifiers, scratchpads, and tool outputs inflate tokens; and credit assignment benefits from replaying or re grounding earlier steps. Today’s models remain constrained by finite context windows and GPU memory, forcing practitioners to truncate histories, compress aggressively, or shard ad hoc—each introducing brittleness, lost signal, or engineering complexity. As tasks become more complex and supervision more process based, context itself becomes the bottleneck resource during training, not just during inference.

We propose a systems research agenda that treats long context as a first class systems problem for RL post training. The core idea is to virtualize and manage context over a memory hierarchy—GPU HBM, host DRAM, and fast storage—while providing compiler/runtime support, algorithms, and telemetry that preserve learning signal with predictable cost. Concretely, this involves KV cache orchestration and eviction policies for multi turn rollouts; context parallelism and sequence aware sharding across accelerators; process memory abstractions for learned summarization and landmarking; verifier aware, process based reward pipelines that do not require full uncompressed histories; and reproducible, fault tolerant checkpointing of stateful training runs. The outcome is a principled stack that lets RL post training operate on long horizon reasoning without prohibitive memory pressure or brittle truncation.

Topics:

Memory management & caching system design that supports infinitely growing context with fault-tolerance and reproducibility
Model-memory interface design for efficient indexing and summarization over growing context during post-training
Exploring algorithmic trade-offs of system optimizations that reduce memory bandwidth pressure during post-training

16. AI-automated HW architecture design

Hardware design is a complicated task that requires many work hours of highly technical personnel with a lot of experience and expertise. AI advancements in recent years present a unique opportunity to automate parts of this process and democratize custom chip creation.

This project aims to study how some workload of computer chip architects can be automated and offloaded using AI. Examples include:

Architecting the right cache configuration
Choosing the right tape-out node
Creating new datatypes
Coming up with new memory technologies
Etc.

Each of these examples is a research project of its own, requiring:

Selecting the right HW representation
Collecting training dataset
Training/fine-tuning AI models
Using these models to find novel HW or HW configurations

Topics:

Computer architecture and HW design
Training/fine-tuning AI models

17. Towards Unified Multimodal Representation

Multimodality research nowadays mainly focuses on image and language, which is far from accommodating all modalities that can be present in our physical world. As different modality exhibits different characteristic and specialty, and the way modalities interact with each other hugely differs, we aim to build up unified multimodal representation that essentially maintains different modality’s specialty while accommodating cross-modality multi-faceted interactions. The unified multimodal representation can be delegated by vision, code, language or latent space.

You will be working on RL-based modality coordination, contrastive learning and modality aware infrastructure design, and LLM-centered multimodality learning and foundational model training.

Research Objective:

Gain in-depth understanding of multimodal intersections.
Investigate how to build up or learn unified multimodal representation.
Top-tier conference paper publication or technical report.

18. Reasoning in the Age of Multi-Agent LLMs

As Large Language Models (LLMs) evolve into multi-agent ecosystems, reasoning—a cornerstone of intelligence—faces a fundamental transformation. Traditionally, reasoning has been viewed as a centralized process within a single model. However, in multi-agent systems, reasoning becomes a distributed capability, emerging from interactions among autonomous agents that share partial knowledge and collaborate toward common goals. This distributed nature introduces new dynamics: reasoning is no longer confined to a single locus but is shaped by communication protocols, coordination strategies, and emergent behaviors across agents.

Understanding this shift is critical for advancing agentic AI. Distributed reasoning raises profound questions: How do agents collectively construct coherent logic? What vulnerabilities arise when reasoning fragments across multiple entities? Current approaches often fail to ensure consistency, robustness, and efficiency in such settings, leading to coordination failures and systemic biases. This research aims to uncover the principles governing reasoning in multi-agent LLM systems, diagnose existing flaws, and design mechanisms—such as shared memory architectures, reasoning protocols, and meta-reasoning layers—that enable reliable and scalable distributed intelligence.

Research Objectives:

Characterize Distributed Reasoning
Investigate the nature of reasoning in multi-agent LLM systems, analyzing how reasoning emerges through collaboration and negotiation among agents.
Identify Systemic Flaws
Expose limitations in current multi-agent reasoning, including inconsistencies, coordination failures, and emergent biases.
Develop Improvement Strategies
Explore architectural and algorithmic approaches to enhance distributed reasoning, such as shared memory mechanisms, reasoning protocols, and meta-reasoning layers.

Impact: This research will establish foundational insights into distributed reasoning, enabling the design of multi-agent systems that are more coherent, resilient, and capable of complex problem-solving. By addressing systemic flaws and proposing scalable solutions, we aim to advance the next generation of agentic AI for collaborative and autonomous tasks.

19. Multi Agent RL for Improving LLMs’ Performance

Multi agent reinforcement learning (MARL) offers a principled pathway to train language models to reason through interaction, feedback, and strategic pressure—rather than relying solely on static supervision or heuristic chain of thought templates. In MARL, populations of complementary agents (e.g., planner, decomposer, solver, verifier, critic) coordinate or compete within task environments that provide verifiable signals of reasoning quality. Rewards can be derived from formal checks, unit tests, simulators, constraint satisfaction, or outcome metrics, turning abstract “good reasoning” into an optimizable objective. Centralized training/decentralized execution (CTDE) and league style curricula enable agents to confront progressively harder counterexamples and adversaries, while population diversity and role specialization encourage robust strategies such as problem decomposition, tool use, and self correction.

Realizing this vision requires careful attention to training signals, optimization stability, and credit assignments. We propose process based rewards that score intermediate reasoning steps (not just final answers), trajectory aware attribution to identify which interactions improved the solution, and opponent/teammate sampling to prevent mode collapse. MARL will be coupled with programmatic evaluators—SAT/SMT solvers, theorem provers, code tests, simulators—to reduce reward misspecification and provide grounded feedback loops. The outcome is an LLM that internalizes transferable reasoning skills—planning, verification, and abstraction—that persist even when deployed as a single agent, yielding higher accuracy, stronger robustness, and safer behavior under uncertainty.

Main Research Questions and Directions:

General MARL Schema: How do we define a general multi agent RL formulation—task decomposition, role taxonomies, communication protocols, and reward structure—that applies across domains and horizons?
Task → Role Mapping: How can we automatically distinguish subtasks (planning, search, calculation, verification, reflection) and assign the right roles at the right time?
Rewarding the Process: What verifier coupled, process based rewards best correlate with genuine reasoning quality (correctness, groundedness, minimal hallucination, tool efficacy)?
Credit Assignment: How do we attribute causal contribution to specific messages, tools, or agents in long, intertwined trajectories without encouraging reward hacking?
Optimization Stability: Which combinations of league training, opponent/teammate shaping, and diversity regularizers yield stable, non collapsing population dynamics?
Meta Control: Can a meta controller learn when to switch roles (plan→solve→verify→reflect), how to escalate, and what to summarize to preserve signal under context limits?
Transfer & Distillation: How do we distill team competencies into strong single agent policies while retaining planning, verification, and abstraction skills?
Generalization & Safety: How do we measure OOD transfer, adversarial robustness, tool use safety, and bounded hallucination under MARL trained policies?

20. Evaluating and Enhancing LLMs for System and Hardware Design

System and hardware design represents one of the most demanding domains in computing, requiring deep technical reasoning, trade-off analysis, and mastery of complex abstractions. While LLMs have demonstrated impressive capabilities in code generation and documentation, their readiness to tackle system-level design challenges remains uncertain. Designing distributed systems or hardware architectures involves balancing performance, power, area, scalability, reliability, and security—tasks that demand structured reasoning and domain-specific expertise beyond pattern matching or text synthesis.

This research seeks to rigorously evaluate the current state of LLMs in system and hardware design and identify gaps that limit their effectiveness. Are LLMs capable of reasoning about architectural trade-offs, protocol optimizations, and resource allocation strategies? Where do they fail—lack of contextual awareness, inability to handle multi-objective optimization, or insufficient grounding in decades of systems research? Building on this assessment, we will develop strategies to enhance LLM contributions, including fine-tuning with system-specific corpora, integrating symbolic reasoning frameworks, and leveraging reinforcement learning with system feedback. The ultimate goal is to unlock LLMs’ potential as active collaborators in designing next-generation computing and hardware systems.

Research Objectives:

Assess Current Capabilities: Benchmark LLMs on representative system and hardware design tasks, such as GPU architectural trade-offs, resource allocation, and protocol design.
Identify Gaps and Limitations: Analyze where LLMs fall short—e.g., lack of contextual reasoning, inability to handle multi-objective optimization, or insufficient domain knowledge.
Enhance Design Competence: Develop methods to improve LLM performance, including fine-tuning with system-specific corpora, integrating symbolic reasoning, and leveraging reinforcement learning with system feedback.

Impact: This work will clarify the role of LLMs in system research and propose pathways to unlock their potential for designing next-generation computing systems. By bridging the gap between AI capabilities and system-level reasoning, we aim to accelerate innovation in distributed systems and hardware architecture.

21. Enhancing Performance in LLM Agents with External Tools

As large-scale LLM-based agents are increasingly deployed in real applications, their ability to collaborate efficiently with external tools—such as web search engines, vector databases, knowledge bases, code execution environments, and even additional LLM inference services—has become critical.

However, the current agent ecosystem remains fragmented: each external tool operates as an independent black box, with no shared execution context, no cross-tool optimization, and no visibility into underlying system performance. This leads to inefficient orchestration, unnecessary latency, and suboptimal resource utilization when agents attempt to combine reasoning with external tool capabilities.

This collaboration aims to develop new methodologies and system frameworks that enable coordinated, observable, and optimized cooperation between LLM agents and external tools. By introducing systematic measurement techniques, building unified orchestration layers, and exploring model–system co-design, we seek to significantly improve the overall execution efficiency of LLM agent pipelines. Our ultimate goal is to reduce end-to-end latency, improve throughput, and enable more capable and scalable agent systems.

Topics:

Bottleneck Identification and Analysis in LLM Agent + External Tool Systems
Optimization Techniques for Coordinated Tool Invocation in Heterogeneous Agent Infrastructures

22. Advancing Scalable AI Systems for Large-Scale Training, Inference, and Intelligent Networked Services

As AI and cloud infrastructures continue to grow in scale and complexity, effectively identifying and addressing system bottlenecks in large-scale distributed training, LLM inference, and AI-enabled networked services has become critically important. These systems integrate diverse components such as GPU clusters, heterogeneous accelerators, neural media pipelines, microservice architectures, and high-performance datacenter networks. However, the rapid evolution of deep learning models—combined with the increasing diversity of hardware and network environments—creates intricate performance challenges that demand systematic and scalable optimization strategies.

This collaboration aims to develop innovative methodologies for measuring, analyzing, and resolving bottlenecks in AI/ML systems, particularly within distributed model training, memory-constrained LLM architectures, edge-assisted large-model serving, and neural media delivery frameworks. By employing precision measurement techniques to diagnose computation, communication, and scheduling inefficiencies—and by applying model-driven and system-driven solutions—we seek to improve resource utilization, reduce tail latency, and enhance the overall efficiency of AI execution pipelines. Our ultimate goal is to enable scalable, reliable, and high-performance AI infrastructures capable of supporting next-generation deep learning, intelligent media processing, and networked AI services.

Topics:

Bottleneck Identification in Large-Scale Distributed Training and LLM Inference Pipelines
Performance Analysis of GPU Clusters and Heterogeneous AI Accelerators
Computation–Communication Co-Design for High-Performance AI
Network-Aware Optimization in Datacenter-Scale AI/ML Infrastructure

23. LLM-Assisted Research in Network Domain

The emergence of Large Language Models (LLMs) presents a transformative opportunity for network research, where complexity and scale often hinder rapid progress. Traditional workflows for network testing and analysis require significant manual effort, slowing innovation and limiting adaptability. By embedding LLMs into these processes, we can automate test generation and evaluation, streamline data interpretation, and enable predictive insights that empower human experts to focus on strategic decisions. This approach not only accelerates research but also ensures reliability, transparency, and compliance in critical network environments, paving the way for intelligent, collaborative systems that address the growing demands of modern networking.

Topics:

Automated Network Testing: LLM-driven generation, execution, and interpretation of network test cases.
Intelligent Data Analysis: Using LLMs for anomaly detection, performance optimization, and trend prediction.
Human-AI Collaboration: Designing workflows that integrate expert judgment with LLM-assisted insights.

24.Toward Future Architectures for LLMs and Agentic Systems

Large Language Models (LLMs) are evolving at an unprecedented pace. Innovations in attention mechanisms, quantization techniques, sparsity methods, and model architectures continue to reshape the landscape. Beyond the core algorithms, the emergence of LLM-based agentic systems introduces additional complexity, as these systems integrate diverse applications such as databases and external tools. Currently, GPUs dominate LLM serving hardware due to their ability to deliver high throughput for large-scale dense matrix multiplications. However, this dominance comes with limitations. GPU efficiency is constrained in scenarios that demand varying latency and energy profiles, making them less optimal for heterogeneous workloads and real-time agentic interactions. What’s the right future architecture for LLMs and LLM-based system remains an open question.

Topics:

Automated LLM accelerator generation: A toolchain to automatically map LLM algorithm to the FPGA hardware
Hardware accelerated agentic system: Accelerate the agentic system with heterogeneous hardware (e.g. FPGA)

– 지원기간: 2026.1.1 – 2026.12.31

– 지원내용: 프로젝트경비 (각 선정 과제별 추후 별도 확인) – 총 9천 – 1억 8천만원

정부부문: 각 선정 과제별 추후 별도 확인 (Korea Won 8천만원 – 1억 7천만원)
기업부문: 각 선정 과제별 추후 별도 확인 (Korea Won 9백 -1천만원)

* 정부부문의 프로젝트비 산정. 사용 등은 정보통신방송 연구개발 관리규정에 따름 (IITP의 추후 안내)

* 기업부문은 기업과제 별도 계약에 따름.

– 선정심사: 마이크로소프트연구소의 전문성심사(서면)

* 심사결과는 선정된 과제에 한하여 개별 통보하며 공개되지 않음

* 과제 공동연구기관으로 선정된 국내 대학은 IITP와 협약 등 정부 과제 수행을 위한 절차를 추진해야 함(대학별로 공동연구기관으로 협약 체결)

B. 공동연구원 파견

– 공동 연구원: 별도 심사를 통해 선발

– 파견기간: 180일 (2026. 3월 – 2026. 9월 중 예정, 파견시점 파견 국가 규정에 따라 파견 기간 변경 가능)

– 파견기관: 마이크로소프트연구소 (연구 분야에 따라 중국 북경, 상해, 캐나다 벤쿠버 기타 지역 등에 파견 예정, 파견지역은 해외파견 지 상황에 따라 일부 변경 가능)

4. 신청자격

프로젝트당 학생(2-5명) 및 지도교수로 팀을 구성

학생: 국내 IT관련학과 대학원에 재학중인 전일제 석박사과정 대학원생

* 한국 국적의 내국인 (휴학생 또는 박사후 과정은 제외)

교수: 국내 IT관련학과 소속 전임교원으로서 지원기간 동안 프로젝트 총괄 및 학생 연구 지도가 가능한 자

5. 지원절차

프로젝트 선정 공고 -> 제안서 제출(온라인 접수, 지원양식, 100% 영문제안) -> 선정심사 -> 지원대상 선정통보 -> 협약체결 및 프로젝트 경비 지급

* 지원양식에서 예산작성은 기업부문 Korea Won: 1천만원 기준으로 작성. 정부부문의 프로젝트비는 선정통보후 별도 안내 예정

6. 신청 유의사항

프로젝트팀은 총1개 분야에 한해 신청할 수 있음. 그러나 특수한 경우, 주관연구개발기관 연구책임자(마이크로소프트연구소, 이미란 전무)의 승인 하에 최대 2개 분야에 신청할 수 있도록 함.

신청자격에 부합하지 않을 경우 선정심사 대상에서 제외될 수 있음

* 국가연구개발사업에 참여제한 중인 자와 기관은 신청 불가

7. 신청요령

신청방법: 이메일 신청([email protected])

신청접수마감: 2026년 1월 10일(토) 17:00

* 제출된 서류는 일제 반환되지 않음

8. 문의처

사업담당: 마이크로소프트연구소 이미란 전무 (010-3600-4226, [email protected])