Volume 45

Lawrence Livermore National Laboratory (LLNL) participated in the first 1,000 Scientist AI Jam Session, bringing together over 1,400 DOE scientists to explore AI's role in scientific discovery. Hosted at nine national labs, the event allowed hands-on testing of advanced AI models for research applications. 

 Brian Spears, director of LLNL’s AI Innovation Incubator, emphasized, "We want to get these models in the hands of staff… and help them understand what they can do with them and what they can't yet do with them." Over 300 LLNL experts worked in teams to test AI’s capabilities in data analysis and problem-solving. 

The event identified AI's strengths and limitations, paving the way for future collaborations. More engagements with AI companies are planned to refine AI-driven tools for hypothesis generation, experiment automation, and coding efficiency in scientific research. Read the full article here.  

LLNL researcher Bhavya Kailkhura co-authored the paper "Low-rank Finetuning for LLMs is Inherently Unfair," which received the Best Paper Award at the AAAI 2025 CoLoRAI Workshop held in Philadelphia in March. This research critically examines Low-rank Adaptation (LoRA), a popular method for fine-tuning large language models (LLMs) due to its computational efficiency. The team’s research reveals that LoRA may inadvertently preserve undesirable biases and toxic behaviors in LLMs, especially when fine-tuning is intended to mitigate such issues. Through comprehensive empirical evidence across various models, datasets, and tasks, we demonstrate that LoRA's limitations in capturing shifts in data distribution can lead to unintended consequences. This can inadvertently amplify AI harms—intensifying bias, unfairness, and creating a false sense of safety in AI systems. From Kailkhura: “This is a call to the AI community to reflect deeply on the unintended consequences of fine-tuning methods that may compromise model fairness. We hope this work prompts careful evaluation of fine-tuning methods to promote responsible LLM development.” Congratulations to his co-authors: Ferdinando Fioretto, Saswat Das, and Marco Romanelli from the University of Virginia; Cuong Tran from Dyania Health, and Zarreen Reza from OpenMined. Read the full paper here. 

LLNL is leading a three-year $7 million initiative in collaboration with Oak Ridge National Laboratory and universities to integrate LLMs into high-performance computing (HPC) software development, aiming to enhance performance and sustainability. Spearheaded by computer scientist Harshitha Menon, the project focuses on creating LLM-powered agents to assist developers with complex tasks. These agents operate in a "Sense-Plan-Act" loop: they analyze code and performance data ("sense"), devise strategies like identifying code hotspots ("plan"), and execute tasks such as code generation ("act"). User interactions, including approvals and feedback, further refine the agents' outputs. The team also emphasizes data quality validation, output verification, and model interpretability to ensure the reliability of LLM-generated code. By demonstrating this approach on the Extreme-scale Scientific Software Stack, developed during the Department of Energy's (COE) Exascale Computing Project, LLNL aims to reduce maintenance costs and promote broader adoption of HPC software tools, aligning with evolving industry advancements. This project is funded by the Advanced Scientific Computing Research program within the DOE’s Office of Science. Read the full article.   

Data scientist Luke Jaffe was selected for a poster presentation at the Winter Conference on Applications of Computer Vision (WACV), the premier international computer vision event. He presented a poster on the topic of person localization in images from his recent paper titled, “Swap Path Network for Robust Person Search Pre-Training,” co-authored with a UC Berkeley colleague. The work was conducted as part of Jaffe’s PhD thesis which he completed last year while working at LLNL on representation learning for efficient localized image retrieval. His research shows that the method still works even where there is a high degree of label noise (i.e., many people in the images are unlabeled, or many extra “false positive” labels do not match people). The method is also useful in that it can be applied to completely unlabeled data to help pre-train models for localizing any object type, not just people—an area of future research for Jaffe, who is profiled below. 

LLNL researchers were also on hand for NVIDIA’s flagship conference, GTC 2025, which showcases the latest in AI, edge computing, and autonomous automation. Derek Mariscal was part of a session entitled “Making Physical AI a Reality Through Real-Time Edge Computing” and Abhik Sarkar presented on “The Instrumental Edge: Enabling Real-Time AI Scientific Discovery.” From Abhik: “NVIDIA GTC is a true testament to the collaborative spirit driving technological advancements across a wide range of industries. From enabling tele-surgery from a coffee shop to leveraging AI for real-time signal processing at terabits-per-second data rates in radio astronomy, these transferrable innovations are a powerful boost to advance our vision of self-driving laboratories. The time to act 'together' is now! 

For centuries, drug discovery was a slow, trial-and-error process, sometimes taking decades to develop life-saving treatments. But what if we could speed up that timeline? At LLNL, scientists are using supercomputing, machine learning, and AI to revolutionize how new medicines are found, tested, and developed. 

This podcast episode takes a deep dive into the cutting-edge technology behind faster drug discovery, how researchers are using high-performance computing to target diseases more precisely, and the potential to bring new treatments—from cancer drugs to viral countermeasures—to patients in record time. 

Listen to the entire Big Ideas Lab episode on Apple or Spotify. 

In a recent AI & Data Exchange 2025 interview, Brian Spears and Judy Hill discussed how LLNL’s new El Capitan supercomputer is driving AI and scientific advancements. El Capitan, 16 times more powerful than its predecessor, will revolutionize national security modeling and HPC. Spears, director of LLNL’s AI Innovation Incubator, emphasized AI’s role in accelerating complex problem-solving, stating, “We’re no longer just running simulations. We’re teaching AI to navigate vast scientific landscapes faster than ever before.” He highlights how AI enhances fusion energy research by optimizing experimental designs. Hill underscored El Capitan’s ability to handle both precision simulations and AI workloads, placing LLNL at the cutting edge of AI-driven science. This milestone marks a new era of HPC, where AI and supercomputing work together to solve national security and global scientific challenges. 

Luke Jaffe is a data scientist in the Global Security Computing Applications Division, where he applies computer vision to a variety of areas such as facial recognition, adversarial machine learning, battery defect detection, and power grid infrastructure for increased resilience. “I’m working on recognizing things at microscopic scale in material defect images, but also at macroscopic scale in images from handheld cameras and satellites,” Jaffe said. “The cool part is that you can use the same modeling techniques at these vast differences in scale.” Jaffe began his time at LLNL as an intern in 2015, joined full-time in 2016, and completed his PhD through the Employee Tuition Assistance Program. Through word of mouth, he often supports and advocates for other employees seeking to make use of this valuable program. This March, he represented DSI at the Winter Conference on Applications of Computer Vision (see story above). “It’s important to have people from Livermore at top computer vision conferences, as it is an overlooked part of AI right now compared to large language models,” he said.