{"id":144931,"date":"2020-02-27T20:34:30","date_gmt":"2015-01-30T00:14:18","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/group\/deep-learning-technology-center\/"},"modified":"2026-10-22T14:09:43","modified_gmt":"2026-10-22T21:09:43","slug":"deep-learning-group","status":"publish","type":"msr-group","link":"https:\/\/www.microsoft.com\/en-us\/research\/group\/deep-learning-group\/","title":{"rendered":"Deep Learning Group"},"content":{"rendered":"
\n\t
\n\t\t
\n\t\t\t\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\tReturn to Microsoft Research Lab – Redmond\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

Deep Learning Group<\/h1>\n\n\n\n

Models that reason, act, and collaborate<\/em><\/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

The Deep Learning Group at Microsoft Research develops the scientific foundations of AI systems that reason, act in the world, and collaborate with people. Our modeling goal is to connect these capabilities: reasoning organizes computation, interaction supplies experience, and collaboration makes human intent part of the task. We build learning methods, models, and research tools that make these connections concrete. Coding, computer use, robotics, tutoring, and data analysis provide distinct settings for studying the same underlying problems of representation, feedback, and adaptation.<\/p>\n\n\n\n

\n\t\n\t
\n\t\t
\n\t\t\t

Research directions<\/h2>\n\n\n\n\n\n

Reasoning requires more than producing a correct final answer. A model must represent intermediate states and learn which computational paths are useful. We study the learning objectives and internal structures that support this process, including the relationship between neural representations and symbolic computation.<\/p>\n\n\n\n

GFlowRL (opens in new tab)<\/span><\/a> uses distribution-matching reinforcement learning to retain alternative reasoning paths. Progressive Thought Encoding (opens in new tab)<\/span><\/a> compresses intermediate reasoning into fixed-size states to bound memory use. Current neural-symbolic research (opens in new tab)<\/span><\/a> identifies tensor-product structure within learned representations and uses it to study targeted changes in model behavior. PlugMem (opens in new tab)<\/span><\/a> connects neural language models with an explicit knowledge graph of facts and action guidance, supporting memory retrieval and reasoning.<\/p>\n\n\n\n\n\n

An agent learns through actions and their consequences. We connect model training to environments where tasks require tools, visual grounding, and decisions over multiple steps. This places the environment and the feedback signal inside the research problem, rather than treating them only as evaluation infrastructure.<\/p>\n\n\n\n

Orchard (opens in new tab)<\/span><\/a> provides an open agent-modeling framework with reusable environments and training and evaluation recipes for coding, web interaction, and assistant tasks.<\/p>\n\n\n\n

Magma (opens in new tab)<\/span><\/a> connects vision-language understanding to grounded actions, using image, video, and robotics data to support interface navigation and robotic manipulation.<\/p>\n\n\n\n

Argos (opens in new tab)<\/span><\/a> uses an adaptive verifier to select scoring tools and connect multimodal training rewards to visual and temporal evidence.<\/p>\n\n\n\n\n\n

In collaboration, people refine goals, provide feedback, and change direction as work develops. We study models and interfaces that support this process, with learning signals that reflect the interaction rather than only the next response.<\/p>\n\n\n\n

CollabLLM<\/a> trains for contributions across multiple turns, rewarding clarification and proactive assistance that advance human\u2013AI collaboration rather than optimizing each reply in isolation.<\/p>\n\n\n\n

EvoLib<\/a> transforms raw experience into an evolving library of knowledge that improves downstream performance consistently over time. StudentSim (opens in new tab)<\/span><\/a> trains individualized student simulators from sparse learner records to support tutoring research.<\/p>\n\n\n\n

Data Formulator (opens in new tab)<\/span><\/a> combines AI assistance with interactive data preparation, exploration, and visualization. Flint (opens in new tab)<\/span><\/a> compiles compact chart specifications using semantic data types. These tools make intent and revision explicit parts of data work.<\/p>\n\n\n\n

Speaking the World Into Existence<\/a> applies recent progress in large models towards rapid creation of virtual scenes, with applications to assisting in game development and beyond to rapid creation of interactive training scenarios.<\/p>\n\n\n\n\n\n

Reliable interaction requires testing sequences of behavior, not just isolated outputs. We study conversational safety and secure coding alongside model capability. The objective is to expose failures in the settings where they arise and provide concrete feedback for model and agent development.<\/p>\n\n\n\n

SEMA<\/a> studies automated multi-turn safety stress testing. MultiBreak (opens in new tab)<\/span><\/a> provides a benchmark for safety evaluation across extended conversations. SecureVibe (opens in new tab)<\/span><\/a> trains coding agents to account for implicit security requirements through planning and testing.<\/p>\t\t<\/div>\n\t<\/div>\n\n\t<\/div>\n\n\n\n

Scientific perspective<\/h2>\n\n\n\n

Our research connects learning objectives, model representations, and interaction protocols. Each shapes what a system can learn and how its behavior can be examined. We develop this science with collaborators through concrete models and public research artifacts, bringing foundational work into contact with the demands of action and collaboration.<\/p>\n\n\n\n

\n
Explore our publications<\/a><\/div>\n<\/div>\n\n\n","protected":false},"excerpt":{"rendered":"

The Deep Learning group advances the state-of-the-art in deep learning\u00a0to achieve general intelligence. We\u00a0develop algorithms, models, and systems\u00a0in deep supervised and unsupervised learning, deep reinforcement learning,\u00a0and neural-symbolic reasoning, and then pursue breakthroughs in computer vision, natural language processing,\u00a0multimodal intelligence, time series analysis, and other relevant areas.<\/p>\n","protected":false},"featured_media":0,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr_group_start":"","footnotes":""},"research-area":[13561,13556,13562,13545],"msr-group-type":[243694],"msr-locale":[268875],"msr-impact-theme":[],"class_list":["post-144931","msr-group","type-msr-group","status-publish","hentry","msr-research-area-algorithms","msr-research-area-artificial-intelligence","msr-research-area-computer-vision","msr-research-area-human-language-technologies","msr-group-type-group","msr-locale-en_us"],"msr_group_start":"","msr_detailed_description":"","msr_further_details":"","msr_research_lab":[199565],"related-researchers":[{"type":"guest","display_name":"Jyoti Aneja","user_id":700582,"people_section":"Section name 0","alias":""},{"type":"user_nicename","display_name":"Harkirat Behl","user_id":41548,"people_section":"Section name 0","alias":"hbehl"},{"type":"guest","display_name":"Yonatan Bisk","user_id":788162,"people_section":"Section name 0","alias":""},{"type":"user_nicename","display_name":"Hao Cheng","user_id":39922,"people_section":"Section name 0","alias":"chehao"},{"type":"user_nicename","display_name":"Michel Galley","user_id":32887,"people_section":"Section name 0","alias":"mgalley"},{"type":"user_nicename","display_name":"Jianfeng Gao","user_id":32246,"people_section":"Section name 0","alias":"jfgao"},{"type":"user_nicename","display_name":"Lars Liden","user_id":32612,"people_section":"Section name 0","alias":"laliden"},{"type":"user_nicename","display_name":"Xiaodong Liu","user_id":34877,"people_section":"Section name 0","alias":"xiaodl"},{"type":"user_nicename","display_name":"Swadheen Shukla","user_id":38248,"people_section":"Section name 0","alias":"swads"},{"type":"user_nicename","display_name":"Paul Smolensky","user_id":36353,"people_section":"Section name 0","alias":"psmo"},{"type":"guest","display_name":"Ryuichi Takanobu","user_id":646683,"people_section":"Section name 0","alias":""},{"type":"user_nicename","display_name":"Andrea Tupini","user_id":40339,"people_section":"Section name 0","alias":"andreatupini"},{"type":"user_nicename","display_name":"Chenglong Wang","user_id":41251,"people_section":"Section name 0","alias":"chenwang"}],"related-publications":[593410,591757,570588,572352,553113,591013,506918,570432,508631,502184,552729,490436,574707,559176,506933,506330,551151,482721,500372,466368,493946,466362,495725,506924,488027,328118,437235,481140,473007,495689,438609,454764,424101,435120,435126,369872,498779,377222,397103,398654,294695,383546,369965,383549,376403,354179,391253,391196,400139,393209,374186,372953,340424,380765,391262,339806,309554,328361,324893,391292,391277,314969,294440,294713,294722,294719,299276,238331,270507,283535,238203,238137,314984,238121,242609,242621,237177,238157,391304,238284,238202,238204,215435,215434,238173,215430,168904,365066,354209,238152,168801,215140,311459,168759,168302,168551,168020,168486,168443,167899,168021,168090,168019,167946,340433,167876,167045,167739,166529,166513,166451,166669,166530,166528,165896,165215,164901,164452,164450,163543,162865,364730,364694,364721,608640,613824,626973,626988,628500,630861,633303,640596,645273,645300,647163,650868,650883,654966,672750,673644,676965,680259,680343,681282,701698,706384,706405,707338,707347,715066,716989,717001,717016,717262,725137,741730,744436,751342,766183,783775,785344,788207,788249,804010,810415,821809,838312,841159,846181,847417,847423,847432,847438,847444,847483,847984,858045,860352,879321,881913,883905,885843,888294,922803,922809,936855,940530,940548,944574,966210,966927,982488,998313,1004331,1004340,1004352,1004367,1004373,1005141,1007064,1030533,1093152,1114974,1126704,1128837,1136383,1141355,1154630,1166856,1172455,1175764],"related-downloads":[],"related-videos":[301163,775504,1132209,777859,777850,775894,775786,775777,775513,689436,775492,775480,775438,775429,755203,738595,689448],"related-projects":[931254,890049,847462,804847,810817,811012,811027,811039,466380,425502,394646,572997,377990,171217],"related-events":[394283,394319],"related-opportunities":[1186465],"related-posts":[416867,444303,456975,754636,1190353,1177589,1176573,1136909,1101105,1008426,989523,965166,895428,851928,823648,488363,728458,717256,715399,696898,657990,648279,632745,626169,610734,592603,578602,573408,564300],"tab-content":[{"id":0,"name":"Highlights","content":"