IBM Research

Fine-tuning large language models on specialized data the traditional way can involve updating billions to trillions of weights. Low-rank adaptation, or #LoRA, offers a shortcut. With LoRA, you change only a tiny subset of the base model’s weights, creating a plug-in module (also called a LoRA!) that gives your model domain-specific expertise at inference time. Like custom bits for a multi-head screwdriver, LoRAs can be swapped in and out of the base model to give it specialized capabilities. Here are several ways IBM Research is innovating with LoRA to make it easier to customize and serve #AI models at scale. https://lnkd.in/e_4G4rV8

IBM Research recently concluded our annual Pat Goldberg Memorial Best Paper Award competition. This competition recognizes a small number of outstanding papers spanning the broad breadth of IBM Research including computer science and AI, mathematical sciences, physical sciences and quantum computing. These papers also reflect strong collaborations with leading research universities. Papers published in 2023 were eligible, and this year, three papers were selected as winners, and four as honorable mentions. These seven papers were published in leading scientific publications: Nature Magazine, Physical Review Letters, the IEEE Journal of Solid State Circuits, and NeurIPS. Topics covered in the papers include a paradigm using AI to discover fundamental scientific laws; advances in superconducting computers; and key advances enabling practical use of quantum computing before we have fully fault-tolerant systems. These awards were established in the memory of Pat Goldberg, an IBM researcher who was instrumental in establishing systematic research disciplines in IBM Research. The competition was named in her honor when she passed away after retiring from IBM. We salute the winners and encourage you to read the papers here: https://lnkd.in/ehnsyKuj

In this week's newsletter, we are building the future of chips in the United States. This week we announce the home of the new NSTC EUV Accelerator, learn more about advancing quantum algorithms, and continue recognizing impact from IBM in AI. Read more for the latest and subscribe here:

This morning, Natcast.org and NY Gov. Kathy Hochul announced that the Albany NanoTech Complex will be the home of the first National Semiconductor Technology Center (NSTC) site, designated by the 2022 CHIPS and Science Act. Based in Albany, New York, and run by NY CREATES, this complex will stand up the NSTC Extreme Ultraviolet (EUV) Accelerator, a new facility that will accelerate cutting-edge semiconductor research in the United States. IBM and other industry leaders have used EUV technology at Albany NanoTech to produce some of the world’s most innovative chip breakthroughs, including the world’s first 2 nm chip. As a proof point for the new EUV Accelerator, IBM Research has already shown the first working proof of designing wafers using the next-generation High NA EUV technology, before a machine has even made it to Albany. This new research provides a pathway for designing chips below the 2 nm level. Read more about the announcement and what is in store for Albany: https://lnkd.in/eHGmTvch

 IBM Research returns to the top of the BIRD benchmark for handling complex database queries! When IBM’s text-to-SQL generator reached the top of the BIRD benchmark in May, the competition intensified as new companies joined the fray. But on Monday, IBM’s Granite-based solution reclaimed the top spot for both execution accuracy and the efficiency of its generated SQL. Our new text-to-SQL Granite 34B model adds two novel, quality control features to boost accuracy and robustness: a post-processing check that ensures the system accurately reproduced the initial input in natural language; and a rule-based component that replaces the top-ranked SQL result with a lower-ranked one if the system detects common error patterns. All this helped to boost our score on BIRD by five percentage points. Read more about our text-to-SQL solution here: https://lnkd.in/e59rDzkJ Michael Glass, Dharmashankar Subramanian, Anika Schumann, & Alfio Massimiliano Gliozzo

In this week's newsletter, we are building agents, brain-like chips, and tiny semiconductors. We take a look at new IBM agents, explore neuromorphic computing, unlock new yield benchmarks for EUV patterning and more. Check out the latest news and subscribe here:

In this week's edition of The Short, we’re diving into the news out of TechXchange about IBM’s new generation of Granite models. We introduce the third generation of Granite, we discuss safer interactions with AI models, and we explore new use cases for cutting-edge AI. Find out more on the latest news and subscribe: https://lnkd.in/e5Vm8S75

To meet the growing demands of AI, IBM research scientists are working on a new generation of computer chips inspired by the most efficient computing machine we know: the human brain. By merging memory and processing, and in some cases storing data in the same way neurons do, these chips can achieve greater performance than existing processors while using less energy. https://lnkd.in/ekStUV7v

On this day in 1956, the IBM Research Laboratory in Zurich, Switzerland officially opened its doors. Today, the Zurich lab is the home to Nobel Prize-winning work, and talented researchers pioneering emerging industries. Learn more here: https://ibm.co/3UmSO2Z Stay tuned for more moments in IBM's research history. #IBMHistory

For most software developers, working through their backlog of issues feels like a never-ending triage. But what if there were tools that could make finding bugs, suggesting fixes, and testing those ideas as easy as submitting an issue on GitHub? At TechXchange today, IBM’s annual developer conference, a team of researchers previewed IBM SWE-Agent 1.0. It’s the first set of software engineering (SWE) agents of its kind, powered by open LLMs, that can autonomously resolve GitHub issues efficiently. It has the potential to dramatically reduce the time developers spend solving problems in their code. Read more about the work here: https://lnkd.in/eZ8bB3gC