Red Hat Research, the Mass Open Cloud (MOC), and IBM Research launched a new collaborative effort with eight US university research projects to advance the National Science Foundation’s National AI Research Resource (NAIRR) Deep Partnership Program. As members of the AI Alliance, the three groups are already major contributors to projects for building, operating, and improving critical components and standards for open source AI infrastructure.
By selecting a small number of projects to work with closely on development and experimentation in the MOC starting in December 2025, the research teams are helping to advance NAIRR pilot efforts and strengthen the foundations for a sustainable national infrastructure that can advance US research AI capabilities. Successful efforts to build open source AI building blocks through this national collaboration will also allow IBM Research and Red Hat to share working results and open experimental data with the wider global software development and systems engineering community.
The new research projects will focus on improving AI for software development (code creation and testing) and exploring novel techniques for greater efficiency, resource optimization, improved reliability, and manageability for AI infrastructure and applications. Researchers receive access to the MOC, including facilitation support, telemetry, and power consumption data as needed; the Red Hat AI open cloud software stack, including Red Hat Enterprise Linux (RHEL) and Open Shift AI for enterprise application development; and open source AI models and tools from IBM Research. The Deep Partnership program is intended to foster significant, sustained collaboration between industry and researchers in academia on projects of mutual interest. The projects will conclude by June 30, 2026. We’re excited to introduce these eight projects and encourage you to follow their progress over the coming months.
AI for software development
Model merging for code LLMs: reasoning fusion and MoE-aware methods
PI: Stacy Patterson, Rensselaer Polytechnic Institute
Model and adapter merging are powerful techniques for integrating task-specific specialist LLMs into a single model, reducing costs and latency while maintaining quality. However, their applicability to software engineering tasks remains underexplored. This project advances two key areas: systematic integration of reasoning-heavy and lightweight models and specialized merging techniques for Mixture-of-Experts architectures, including expert alignment and routing-optimized fusion.
Evaluating and improving applications of Large Language Models to automated software testing
PI: Alessandro Orso, University of Georgia
LLMs can generate software tests with coverage comparable to traditional methods while producing more readable tests that reflect developer intent. Yet significant gaps remain between developer-written tests and both traditional and LLM-based automated test generation (ATG) tools. This project aims to improve our understanding of the differences between generated and developer-written tests and move ATG tools toward producing more meaningful, maintainable, and developer-aligned test suites.
Techniques for improving AI applications
Adaptive KV cache compression for agentic AI
PI: Mohammad Mohammadi Amiri, Rensselaer Polytechnic Institute
Transformer-based LLMs are increasingly deployed in long-context and agentic settings, where memory efficiency becomes a critical bottleneck. A key challenge lies in the growth of the key-value (KV) cache, which scales linearly with sequence length and strains memory resources during autoregressive generation. This project introduces an adaptive compression framework for the KV cache that dynamically learns compact representations of token information in real time.
Efficient memory offloading for cost- and energy-efficient foundation model training
PI: Nam Sung Kim, University of Illinois at Urbana-Champaign
Foundation model training is often limited by GPU memory, with multi-GPU distribution used as the default workaround. This project proposes transparent memory offloading from GPUs to adjacent CPU memory over PCIe as a cost- and energy-efficient alternative. Using MOC infrastructure, this project will study large-scale workloads, including advanced parallelism and Mixture-of-Experts models, with the aim of guiding future system designs that better integrate GPU–CPU memory hierarchies, lowering infrastructure costs and broadening access to large-model training.
Building reliability and transaction semantics for LLM agents
PI: Indranil Gupta, University of Illinois at Urbana-Champaign
Agents based on LLMs are unreliable and often take actions that leave their shared operating environment in an inconsistent state. This project aims to improve agentic reliability for a wide variety of settings via a series of new techniques that always assure the state of the world (e.g., of the codebase, for coding agents) is correct, even if agents execute erroneous actions.
CLEARBOX: interpreting and improving multimodal LLMs
PI: Deepti Ghadiyaram, Boston University
CLEARBOX is an open source suite of tools for white-box and black-box interpretation and intervention of foundational Multimodal Large Language Models (MLLMs). The toolkit will provide researchers with critical infrastructure to dissect internal representations, diagnose cross-modal alignment failures, thereby developing and deploying safer and more robust MLLMs.
Multimodal semantic routing for vLLM
PI: Junchen Jiang, University of Chicago
The vLLM Semantic Router is an open source framework for directing user prompts to the most appropriate LLM. At present, routing is limited to text-only requests, but emerging applications increasingly involve speech, voice, video, and other modalities. This project proposes to extend the Semantic Router with multimodal routing capabilities, focusing on speech-text pipelines and exploring vision capabilities.
Time Series Data Agent: an agentic system with foundation models for multimodal data
PI: Agung Julius, Rensselaer Polytechnic Institute
This project’s objective is to transform the previously developed Abstract Shape Token-Editing Predictive (ASTEP) framework from a passive analysis tool to an active, agentic system capable of multiple time-series tasks, including optimal control. This will lead to an overarching TS-Agent, an agentic framework that orchestrates the model’s functionalities.
Learn more about these projects on the AI Alliance-NAIRR webpage, explore the NAIRR Deep Partnership program at the NAIRR Pilot website, and watch this space and the Red Hat Research Quarterly for updates from individual projects.
