IBM Research

How do you know the agent you’ve built is running as intended? When you look under the hood of a running car, a mechanic knows what should be happening—it’s no different with agents. Enter AgentOps: a new feature implemented across core IBM solutions like Instana, Concert and watsonx Orchestrate that gives developers deeper visibility into how agentic systems operate, make decisions, and use tools: https://ibm.co/6046BEcE6

Happy birthday, Docling! 🥳🍾🎂 Docling has been a runaway success, rapidly becoming the standard in open source AI-driven document conversion. So much of the world's knowledge is contained in documents—with page formatting, charts, figures, and tables that were designed for human eyes, but which machines have traditionally struggled to understand. Docling unlocks the latent potential of the vast troves of document data that every business and institution has stored away, making it accessible to AI and data analytics. I believe that Docling is an exemplar for a powerful but subtle emerging pattern in AI—a hybrid of models and software that harmoniously and unassumingly blend together to get a job done. The latest model powering Docling is a sophisticated, lightweight VLM, but it almost disappears into the library. It's fashionable to dress everything in loud "agent" clothing these days, but Docling is first and foremost a library that does a job, does it well, and stays out of your way. While the open source project itself is only one year old, it is built on a foundation of strong work over many years by a focused and extremely talented team in IBM Research. There's an exciting roadmap of new features and innovation ahead, and I look forward to many future birthday celebrations!🦆

Synthetic data is instrumental to training and scaling AI without exposing real information, but tracking its misuse is challenging. To address this, IBM researcher Pin-Yu Chen presents TabWak—a new watermarking technique designed to help enterprises safeguard their synthetic tabular data. _____ Explore how we extended multimodal watermarking into the process of generating tabular data here: https://ibm.co/6046BETuK Full paper and methodology here: https://ibm.co/6046BETuK

From the ground breaking ceremony of the first leading edge logic Fab in Japan on September 1st, 2023 to the first 2nm Nanosheet prototype wafers produced with electrical results on July 18, 2025! First 2nm lot from start to electrical test in less than two weeks! ALL world records. We at IBM are truly honored to participate in this ambitious project of the nation of Japan and witness a historic moment in the world. Congratulations to all my colleagues at Rapidus Corporation on achieving such an incredible milestone and many IBMers and partners who contributed to this journey together. https://lnkd.in/gTYiEZNd IBM Research

This week at IBM Research we introduce a new benchmark, AssetOpsBench, that puts industrial agents to the test. The rise of LLM agents has led to a new category of benchmarks, with IBM’s AssetOpsBench being the first-of-its-kind specifically designed for asset management. We also announced a new published paper between Moderna and IBM, highlighting the progress of quantum computing and how it could improve the design of mRNA-based medicine. In this case study, IBM and Moderna reveal how new algorithms advance the ability of today's near-term quantum computers to address specific challenges. Finally, we’re learning from IBM’s Pin-Yu Chen on how an invisible watermark can keep tabs on tabular data for AI. For more on the latest news, subscribe now for updates in AI, quantum computing, semiconductors, and more, from across IBM Research 🔬

In less than three years since its founding, Rapidus has achieved its IIM-1 target milestones – from initial groundbreaking in September 2023, clean room completion in 2024 and, in June 2025, the connection of more than 200 of the world’s most advanced pieces of semiconductor equipment. Today, we’re proud to add another very important milestone to this list with the prototyping of leading-edge 2nm GAA transistors at our foundry, with wafers starting to attain electrical characteristics. Full details here: https://lnkd.in/gkUmgCY9

Excited to be at #ICML2025 to talk Red-Teaming, AI Safety & Security, open-source tools, hackathons and secure AI development at IBM Research! Watch this space 🔴 - Kieran Fraser, PhD will be presenting MAD-MAX at the Data in Generative Models Workshop on July 19 - scalable, automated red-teaming for LLMs. Workshop: https://lnkd.in/dSyjCkvQ Paper: https://lnkd.in/dJEnhCpJ - Also check out CoP, recent work by Pin-Yu Chen on controllable prompt tuning for effective red-teaming. Paper: https://lnkd.in/d2mMbFEk #RedTeaming #AISecurity #AISafety #ICML #ibmresearch

Moderna and IBM have teamed up to explore how quantum computing can accelerate success of future mRNA-based medicines. This research taps into quantum’s potential ability to solve complex optimization problems that classical systems might struggle with. https://ibm.co/6047BEMoL Together, we’ve taken a significant step in applying quantum algorithms to one of the most complex challenges in mRNA medicine—predicting a molecule’s secondary structure. These structures influence how stable a treatment is, how efficiently it produces therapeutic proteins, and how it interacts with the body’s cellular machinery. Solving this problem requires navigating an enormous optimization space beyond the reach of classical computing alone. The research teams at IBM and Moderna demonstrated that techniques like variational quantum algorithms, improved with Conditional Value at Risk (CVaR), can already deliver meaningful performance on present-day quantum hardware. Last year, our teams achieved one of the largest mRNA secondary structure simulations ever run on a quantum processor, leveraging up to 80 qubits and 60-nucleotide sequences. Moderna is already integrating quantum into its research today, embracing the hybrid quantum-classical workflows that we believe will ultimately lead to demonstrable advantage, and set the stage for accelerated drug discovery. Head to the case study linked above for more.

𝗦𝘂𝗽𝗲𝗿 𝘁𝗵𝗿𝗶𝗹𝗹𝗲𝗱 𝘁𝗼 𝘀𝗵𝗮𝗿𝗲 𝘁𝗵𝗮𝘁 IBM Research 𝗶𝘀 𝘁𝗵𝗲 𝗽𝗿𝗼𝘂𝗱 𝗿𝗲𝗰𝗶𝗽𝗶𝗲𝗻𝘁 𝗼𝗳 𝘁𝗵𝗲 #𝗜𝗡𝗙𝗢𝗥𝗠𝗦 𝗥𝗲𝘃𝗲𝗻𝘂𝗲 𝗠𝗮𝗻𝗮𝗴𝗲𝗺𝗲𝗻𝘁 & 𝗣𝗿𝗶𝗰𝗶𝗻𝗴 𝗣𝗿𝗮𝗰𝘁𝗶𝗰𝗲 𝗔𝘄𝗮𝗿𝗱 𝟮𝟬𝟮𝟱! 🏆🎉🎊🍾 Our winning work on the 𝗖𝗼𝘂𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝘁𝘂𝗮𝗹 𝗢𝗽𝘁𝗶𝗺𝗮𝗹 𝗔𝗰𝘁𝗶𝗼𝗻 𝗧𝗿𝗲𝗲 (𝗖𝗢𝗔𝗧) for airline ancillary pricing combines causal ML with large-scale mixed-integer optimization. In a live deployment with a global carrier, COAT delivered a ~7% uplift in upsell revenue, with a projected $100M in annual revenue gain from full domestic rollout. It’s not every day that a project, formulated and implemented entirely within industry, receives such meaningful recognition from the academic community. I’m deeply grateful to the award committee for the opportunity to showcase our work. ❤️ https://shorturl.at/3FOU3 That said, this milestone is part of a broader, multi-year journey to integrate cutting-edge AI with the mathematical rigor of OR. We’ve collaborated extensively with academic partners and benefited immensely from their insights, both in breadth and depth, which have helped accelerate and elevate our research. Beyond pricing, 𝗰𝗮𝘂𝘀𝗮𝗹 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻-𝗺𝗮𝗸𝗶𝗻𝗴 has found its way to new frontiers, including LLM routing, to help build more sustainable and efficient AI systems https://shorturl.at/ubgT2. I’m especially grateful to the incredible team behind this work - Shiva S., Youssef Drissi, Zhengliang Xue, Markus Ettl. We couldn’t have done it without each of you. A special shoutout to Markus, my manager for six years - thank you for your trust, for giving us autonomy, and for fostering a space for deep thinking and creativity, which is critical for true innovation. As an OR PhD, I’d also like to dedicate this award to 𝗥𝗮𝗹𝗽𝗵 𝗚𝗼𝗺𝗼𝗿𝘆. While many in the OR community know him as a pioneer in integer programming, fewer may realize he led IBM Research for nearly two decades https://lnkd.in/ePB5tXsw He championed bold, practical innovation, a philosophy that remains core to IBM Research today. In an era where AI is transforming every industry, we’re proud to carry forward that legacy, demonstrating how the synergy of AI and OR can deliver intelligent systems that are not only innovative, but efficient, reliable, and high-performing at scale. Go #OR! Go #IBMResearch! 🫶 #OperationsResearch #Optimization #CausalML #SustainableAI #IBM

Our new IBM Granite 3.3 2B vision model recently debuted at number two on the OCRBench leaderboard, and is the most performant multimodal model under 7B parameters. In enterprise workflows, a massive share of critical information lives in visual formats: charts, tables, scanned documents, diagrams. Text models ignore these. Multimodal models struggle with them. But IBM’s new open-source Granite Vision 3.3 is starting to crack the code. 𝗪𝗵𝗮𝘁 𝗶𝘀 𝗚𝗿𝗮𝗻𝗶𝘁𝗲 𝗩𝗶𝘀𝗶𝗼𝗻? A compact 2B-parameter vision-language model built on IBM’s Granite LLM. Trained on: - 13.7 million enterprise PDF pages - 4.2 million natural images - 80+ million question-answer pairs on document content It’s tuned to handle real-world mess: invoices, receipts, handwritten notes, charts embedded in reports. 𝗪𝗵𝘆 𝗱𝗼𝗲𝘀 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿? Granite Vision just ranked #2 on OCRBench, beating models from OpenAI and Google. It excels in: - Interpreting charts and tables - Understanding document layouts - Extracting structured data from complex visuals It’s not just OCR. It’s visual reasoning. Enterprise AI needs more than text understanding. It needs document intelligence: - Reading beyond words - Recognizing patterns in dashboards - Summarizing visual data with precision Granite Vision demonstrates that this is possible, and it achieves this at a low cost, with open models. - Download the model on HuggingFace: https://lnkd.in/gKwcRHqe - View the full OCRBench: https://lnkd.in/gFZGbVcG