{"id":1142643,"date":"2025-11-26T06:22:06","date_gmt":"2025-11-26T14:22:06","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/"},"modified":"2026-03-18T09:04:07","modified_gmt":"2026-03-18T16:04:07","slug":"machine-intelligence","status":"publish","type":"msr-group","link":"https:\/\/www.microsoft.com\/en-us\/research\/theme\/machine-intelligence\/","title":{"rendered":"Machine Intelligence"},"content":{"rendered":"
\n\t
\n\t\t
\n\t\t\t\"Colour\t\t<\/div>\n\t\t\n\t\t
\n\t\t\t\n\t\t\t
\n\t\t\t\t\n\t\t\t\t
\n\t\t\t\t\t\n\t\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\n\t\t\t\t\t\t\t\t\t<\/span>\n\t\t\t\t\t\t\t\t\tMicrosoft Research Cambridge\t\t\t\t\t\t\t\t<\/a>\n\t\t\t\t\t\t\t\n\t\t\t\t\t\t\t\n\n

Machine Intelligence<\/h1>\n\n\n\n

Advanced machine learning research, grounded in trust, efficiency, capability.<\/p>\n\n\t\t\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t<\/div>\n\t\t<\/div>\n\t<\/div>\n<\/section>\n\n\n\n\n\n\n\n\n\n\n\n\n\n

The Machine Intelligence team at MSR Cambridge (UK) is dedicated to foundational machine-learning research, guided by the principles of responsible AI, collaboration, and scientific excellence. Our work is grounded in trust, capability, and efficiency, and we are deeply engaged in collaborations that build on these foundations for systems, the sciences, and human-centred AI.<\/p>\n\n\n\n

Our workstreams<\/h2>\n\n\n\n

Cognition<\/strong>:  We are advancing the reasoning capabilities of generative AI through principled machine learning approaches that combine NLP, formal methods, logic, and statistics. By uniting experts across disciplines, we aim to build AI systems that reason reliably, generalize effectively, and operate safely in high-stakes settings. Our work balances theoretical insight with empirical validation to ensure robustness, interpretability, and alignment with human values. <\/p>\n\n\n\n

Memory<\/strong>:  We are developing models of memory that make factual knowledge in large language models transparent, controllable, and fully traceable. Our approach enables precise, provenance-aware knowledge infusion with dynamic editing and access control, allowing models to distinguish grounded from sourceless outputs. In collaboration with Microsoft Research and Copilot teams, we bridge fundamental research and real-world deployment, advancing both scientific understanding and product impact. <\/p>\n\n\n\n

Efficient AI<\/strong>:  We are developing more efficient AI systems by automating quantization through a compiler-based approach grounded in programming language theory, enabling scalable and energy-efficient on-device inference. Our technology powers real-world applications like Copilot+ PCs (via Phi Silica) and the AI Toolkit in Visual Studio Code. In parallel, we are advancing quantization methods and designing next-generation optimizers based on first-principles analyses of learning dynamics to achieve more principled, efficient model training. <\/p>\n\n\n\n

Diffusion<\/strong>:  Diffusion Language Models (DLMs) are a promising alternative to autoregressive models, offering potentially higher generative quality but facing challenges with scalability and efficiency. Building on prior image diffusion research using Fourier domain analysis, we aim to enhance DLM training and inference by applying similar statistical insights to language. <\/p>\n\n\n\n

<\/div>\n\n\n\n\n\n

Cognition<\/h2>\n\n\n\n

The recent surge in Generative AI has revolutionized the creation of new systems and tools, transforming the way we work and live. Evaluating and improving the reasoning abilities of Generative AI is key to understanding their generalization abilities and safely deploy them in critical scenarios.<\/p>\n\n\n\n

We are developing state-of-the art technologies to evaluate and\u202fimprove the reasoning abilities of AI systems\u202fby focusing on principled machine learning approaches. This effort that requires a diversity of skills: Natural Language Processing, Formal Methods, Mathematical Logic, Machine Learning, Statistics, etc. We believe that advancing the reasoning capabilities of AI requires not only technical innovation but also interdisciplinary collaboration and rigorous evaluation.<\/p>\n\n\n\n

By bringing together experts from diverse domains, we aim to build AI systems that can reason more reliably, generalize across tasks, and operate safely in high-stakes environments. Our work is grounded in both theoretical insights and empirical validation, ensuring that the systems we develop are robust, interpretable, and aligned with human values.<\/p>\n\n\n\n

Learn more:<\/h3>\n\n\n\n

A Ladder of Reasoning: Testing the power of imagination in LLMs<\/a>
MSR Blog | August 2025 <\/p>\n\n\n\n

Reasoning Elicitation in Language Models via Counterfactual Feedback<\/a> 
Publication | March 2025 <\/p>\n\n\n\n

Re-Imagine: Symbolic Benchmark Synthesis for Reasoning Evaluation<\/a> 
Publication | March 2025 <\/p>\n\n\n\n

Does Reasoning Emerge? Examining the Probabilities of Causation in Large Language Models<\/a> 
Publication | August 2024 <\/p>\n\n\n\n

<\/div>\n\n\n\n\n\n

Memory<\/h2>\n\n\n\n

We are working on models of memory<\/strong> to make factual knowledge in large language models both transparent and controllable. The goal is to enable high precision knowledge infusion at scale \u2013 with full provenance and access control.<\/p>\n\n\n\n

Our approach makes LLM memory interpretable<\/strong>, with clear source attribution and the ability to detect and mitigate hallucinations by distinguishing between grounded and sourceless outputs. The knowledge made available in the LLM\u2019s memory is also fully manageable<\/strong>, enabling dynamic editing and control of what information is available to the model at runtime.<\/p>\n\n\n\n

We collaborate closely with research science teams across Microsoft Research and applied science teams within Copilot product groups to drive real-world impact. Our work bridges fundamental research and product deployment, contributing to both the scientific community and Microsoft\u2019s Copilot experiences. We publish at leading conferences, such as ICLR and EMNLP, and open source our research to advance the broader field.<\/p>\n\n\n\n

Learn more:<\/h3>\n\n\n\n

KBLaM: Knowledge Base augmented Language Model
Paper<\/a> | Repo (opens in new tab)<\/span><\/a> | Blog<\/a><\/p>\n\n\n\n

Learning to Extract Structured Entities Using Language Models
Paper<\/a> | Repo (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

DiSK: A Diffusion Model for Structured Knowledge
Paper<\/a><\/p>\n\n\n\n

<\/div>\n\n\n\n\n\n

Efficient AI<\/h2>\n\n\n\n

We are building more efficient AI systems by automating quantization<\/strong> through a novel compiler-based approach grounded in programming language (PL) techniques. This automation enables scalable, energy-efficient deployment of models for on-device inference. In parallel, we are advancing the state of the art in quantization itself, with methods such as QuaRot and ongoing research into new compression strategies.<\/p>\n\n\n\n

Our work is already powering real-world applications: our quantization pipeline is used in Copilot+ PCs via Phi Silica, and multiple models are now integrated into the AI Toolkit in Visual Studio Code. By combining automation with cutting-edge quantization, we aim to make high-performance AI more accessible and sustainable.<\/p>\n\n\n\n

We are also designing next-generation optimizers<\/strong>, through rigorous, first-principles investigation of model training and its intrinsic properties; we analyse learning dynamics to rigorously characterize model behaviour and move beyond ad hoc, empirical trial-and-error approaches.<\/p>\n\n\n\n

By integrating theoretical frameworks from optimization theory, dynamical systems, and statistical learning theory, we seek to derive principled enhancements in the training of large-scale models.<\/p>\n\n\n\n

Learn more:<\/h3>\n\n\n\n

Towards Efficient Optimizer Design for LLM via Structured Fisher Approximation with a Low-Rank Extension<\/a>
Publication | February 2025<\/p>\n\n\n\n

SWAN: SGD with Normalization and Whitening Enables Stateless LLM Training<\/a>
Publication | February 2025<\/p>\n\n\n\n

Gradient Multi-Normalization for Stateless and Scalable LLM Training<\/a>
Publication | February 2025<\/p>\n\n\n\n

QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs (opens in new tab)<\/span><\/a>
Publication | December 2024<\/p>\n\n\n\n

Low-Rank Correction for Quantized LLMs (opens in new tab)<\/span><\/a>
Publication | December 2024<\/p>\n\n\n\n

[2412.08585] TurboAttention: Efficient Attention Approximation For High Throughputs LLMs (opens in new tab)<\/span><\/a>
Publication | December 2024<\/p>\n\n\n\n

Pyramid Vector Quantization for LLMs (opens in new tab)<\/span><\/a>
Publication | October 2024<\/p>\n\n\n\n

<\/div>\n\n\n\n\n\n

Diffusion<\/h2>\n\n\n\n

Diffusion Language Models (DLMs) explores the potential of diffusion-based models to exceed the generative quality of traditional autoregressive language models, while overcoming their current limitations in training and inference efficiency. Although early results indicate that DLMs can produce more coherent and diverse outputs than autoregressive models, these benefits have yet to be demonstrated at scale\u2014possibly due to slower training and expensive inference. Our DLM work builds upon our recent image diffusion work where we studied the effect of the forward process in the Fourier domain, allowing us to precisely control image reconstruction across frequencies by considering the signal-to-noise of image spatial statistics. By considering similar ordered statistics in language, we hope to improve training and inference regimes in DLMs.<\/p>\n\n\n\n

Learn more:<\/h3>\n\n\n\n

A Fourier Space Perspective on Diffusion Models<\/a> 
Publication | May 2025<\/p>\n\n\n\n

<\/div>\n\n\n\n\n\n

<\/p>\n","protected":false},"excerpt":{"rendered":"

Advanced machine learning research, grounded in trust, efficiency, capability. The Machine Intelligence team at MSR Cambridge (UK) is dedicated to foundational machine-learning research, guided by the principles of responsible AI, collaboration, and scientific excellence. Our work is grounded in trust, capability, and efficiency, and we are deeply engaged in collaborations that build on these foundations […]<\/p>\n","protected":false},"featured_media":1142843,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr_group_start":"","footnotes":""},"research-area":[13556,13555],"msr-group-type":[243688],"msr-locale":[268875],"msr-impact-theme":[],"class_list":["post-1142643","msr-group","type-msr-group","status-publish","has-post-thumbnail","hentry","msr-research-area-artificial-intelligence","msr-research-area-search-information-retrieval","msr-group-type-theme","msr-locale-en_us"],"msr_group_start":"","msr_detailed_description":"","msr_further_details":"","msr_hero_images":[],"msr_research_lab":[199561],"related-researchers":[{"type":"user_nicename","display_name":"Fabian Falck","user_id":44067,"people_section":"Section name 0","alias":"fabianfalck"},{"type":"user_nicename","display_name":"Wenbo Gong","user_id":42873,"people_section":"Section name 0","alias":"wenbogong"},{"type":"user_nicename","display_name":"James Hensman","user_id":43404,"people_section":"Section name 0","alias":"jameshensman"},{"type":"user_nicename","display_name":"Sushrut Karmalkar","user_id":43674,"people_section":"Section name 0","alias":"skarmalkar"},{"type":"user_nicename","display_name":"Rachel Lawrence","user_id":44130,"people_section":"Section name 0","alias":"ralawrence"},{"type":"user_nicename","display_name":"Chao Ma","user_id":42870,"people_section":"Section name 0","alias":"chaoma"},{"type":"guest","display_name":"Liana Mikaelyan","user_id":1133703,"people_section":"Section name 0","alias":""},{"type":"user_nicename","display_name":"Aditya Nori","user_id":30829,"people_section":"Section name 0","alias":"adityan"},{"type":"guest","display_name":"Mathew Salvaris","user_id":1133699,"people_section":"Section name 0","alias":""},{"type":"user_nicename","display_name":"Amit Sharma","user_id":30997,"people_section":"Section name 0","alias":"amshar"},{"type":"guest","display_name":"Xi Wang","user_id":1133697,"people_section":"Section name 0","alias":""},{"type":"user_nicename","display_name":"Javier Zazo","user_id":41341,"people_section":"Section name 0","alias":"javierzazo"}],"related-publications":[996492,1005534,1091778,1128213,1136012,1140599,1142712,1142730,1142735,1142738,1161894],"related-downloads":[],"related-videos":[],"related-projects":[580699],"related-events":[],"related-opportunities":[],"related-posts":[1133691],"tab-content":[],"msr_impact_theme":[],"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-group\/1142643","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-group"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-group"}],"version-history":[{"count":39,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-group\/1142643\/revisions"}],"predecessor-version":[{"id":1156698,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-group\/1142643\/revisions\/1156698"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media\/1142843"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1142643"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1142643"},{"taxonomy":"msr-group-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-group-type?post=1142643"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1142643"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=1142643"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}