Department of Computer Science and Engineering
Room 402, School of IT Building
Indian Institute of Technology Delhi
Hauz Khas, New Delhi, 110016, India
Email: mausam AT cse DOT iitd DOT ac DOT in,
Phone : +91-11-2659-6076 (O)
Yardi School of Artificial Intelligence (Yardi ScAI)
Indian Institute of Technology Delhi
Hauz Khas, New Delhi, 110016, India
Department of Computer Science and Engineering
University of Washington
Seattle, Washington, 98195, U.S.A.
Email: mausam AT cs DOT washington DOT edu
Phone : +1-206-979-7038 (C)
Starting 21st July 2025, I am back in India after a year long sabbatical at Bloomberg's Artificial Intelligence Group. Given my long absence from India and my own research group, I am currently focused on ongoing activities that require my immediate attention. Consequently, I am unable to engage with external entities on any new initiatives. I apologize in advance for not responding personally to each such email message. The only exceptions can be if a proposed project directly builds on some specific research contributions made by my group in the last couple of years.
Moreover, as always, I am unable to take on winter/summer interns or part-time interns. I am also unable to advise non-IITD undergraduate or Masters students in research projects/theses. Please do not contact me with such requests. I do not have bandwidth to respond to them personally. The only exception is at least a year long, full-time research assistant position.
Bio
Mausam is a Professor of Computer Science at IIT Delhi, and served as the founding head of Yardi School of Artificial Intelligence until September 2023. He is also an affiliate professor at University of Washington, Seattle. Recently, he spent a year long sabbatical working as a Visiting NLP Researcher at Bloomberg's AI group. With an over twenty year research experience in artificial intelligence, he has, over time, contributed to many research areas such as large scale information extraction over the Web, AI approaches for optimizing crowdsourced workflows, and probabilistic planning algorithms. More recently, his research is exploring neuro-symbolic machine learning, computer vision for radiology, NLP for robotics, multilingual NLP, and several threads in intelligent information systems that include information extraction, knowledge base completion, question answering, and dialogue systems. He has over 100 archival papers to his credit, along with a book, two best paper awards, and one test of time award. Mausam was awarded the AAAI Fellow status in 2024 for his sustained contributions to the field of artificial intelligence and unusual distinction in the profession. He was also elected as a fellow of Indian National Academy of Engineering (INAE) -- an apex body for India's most distinguished engineers, engineer-scientists and technologists. He was named as one of the 30 Indian Minds Leading the AI Revolution by Forbes India. He has had the privilege of being a program chair for two top conferences, AAAI 2021, and ICAPS 2017, and also served as an Editor-in-Chief of ACL Rolling Review from 2022 to 2024. He received his PhD from University of Washington in 2007 and a B.Tech. from IIT Delhi in 2001.
Newsworthy
In Feb'26,
concurrent to India AI Impact Summit, I participated in many TV programs: a pre-summit discussion, a
special
broadcast on AI, a conversation on AI and education, and a post-summit discussion (in
Hindi). I also gave bites to this Hindu article on math behind AI.
In Feb'26, I received the Nick McKeown chair professorship by IIT Delhi. In Dec'25
I received the Outstanding Scientist Award by National Academy of Artificial
Intelligence, awarded at NAAI Asia AI Conference, Beijing China.
In Aug'25 I was named one of
India's 100 Most Influential People in AI by Analytics India Magazine.
In Apr'25 I was named one of the 30
Indian Minds Leading the AI Revolution by Forbes India and Accel.
My profile
published in Forbes magazine.
In Nov'24 I became a fellow of Indian National Academy of Engineering, for the
specialization of Computer Engineering & Information Technology.
In Jun'24 I was interviewed by AI
Hub in their meet with AAAI Fellows series. I discuss my career,
information extraction, mentorship and creativity.
Humbled and honored to be elected as a AAAI Fellow, for my contributions to NLP, Planning and
Human Computation, as well as my leadership in AI.
The induction ceremony was at the Fellows Dinner at AAAI 2024.
A bit overwhelmed, as
this is one of most prestigious international honors
an AI researcher can get. Thank
you to all mentors, collaborators, and students over the years.
My NPTEL (public) undergraduate (MOOC) course on artificial intelligence reran starting Jan'24. I have been teaching it every year and this offering had over 55,000 students.
In Dec'23 I appeared on an informal podcast on the future of education.
Had excellent results at EMNLP 2023 -- got four submissions accepted on (1) can GPT4 solve IIT-JEE problems?, (2) how to reduce LLM cost without compromising quality?, (3) low-resource language adapters, and (4) neuro-symbolic temporal knowledge-base completion.
Elated to share that this ACL 2023 accepted all six of our submissions (including one in ACL Findings). The six papers are on various aspects of intelligent information systems, including information extraction, knowledge base completion, question answering, and dialog systems.
Excited to share my first CVPR paper on automatic detection and counting of different types of cells found in histopathology biopsies, useful for predicting Celiac disease. Collaboration with gastroenterologists at AIIMS, New Delhi.
In Mar/Apr'23 I gave bites to three articles: Times of India article and another Times of India article, both on using ChatGPT in education, and a third Silicon India article on generative AI in Indian languages.
Honored to receive the ACL'22 test of time award for OLLIE, which contributed an approach for extracting information in an open-domain way, without supervised training data. Here are my acceptance comments when receiving the award in Dublin, Ireland.
In May'22 I appeared on Dostcast and did an informal podcast on AI in mixed Hindi-English.
Excited to become an editor-in-chief at ACL Rolling Review. Feel free to send me an email to help improve ARR.
Had excellent results at ACL 2022 -- got all three of our submissions accepted. The three papers are on multilingual information extraction.
In Mar'21 I gave bites to a Times of India article on natural language processing.
Proud to be a program co-chair of AAAI 2021 with Kevin Leyton-Brown.
ArnetMiner believes that I am the 56th most influential NLP scholar and 64th most influential AI scholar for the decade 2010-2020. Not sure I deserve to be in this list of greats, but happy to receive the honor, nonetheless!
In Oct'20 I was one of the panelists for Doordarshan program on AI and roadmap.
In Oct'20 I was one of the panelists for a Rajya Sabha TV program on AI for social empowerment.
In Mar'20 I gave bites to a Mail Today article on AI education in India.
In Jan'20 I was one of the panelists for a Rajya Sabha TV program on regulating AI.
In Oct'19 I was one of the panelists for a Rajya Sabha TV program on AI.
In Jun'19 I gave bites to an Economics Times article on AI and India.
In Jun'19 I was one of the panelists for a DD Science program on AI in Hindi.
In Feb'19, I received the Jai Gupta Chair fellowship by IIT Delhi.
In Feb'19 I was interviewed by Factor Daily. The transcript of the interview.
In Oct'18 I was honored to participate in a Niti Aayog panel on AI in front of esteemed audience comprising the PM, council of ministers, heads of PSUs and senior bureaucrats in the Govt of India.
In Oct'18 I was interviewed as one of the experts for a Lok Sabha TV program on AI (in Hindi).
In Sep'18 I was interviewed as one of the experts for Rajya Sabha TV features on AI: the video in English and the video in Hindi.
In Aug'18 I gave bites to an India Today article on the future of AI.
In Jan'18 I recorded a public talk on Artificial Intelligence: Past, Present and Future and a Student Q&A session for Living Science.
In Jun'17 I was a Program co-chair for the 27th International Conference on Automated Planning and Scheduling in Pittsburgh.
In Jul'16 I was invited to deliver a talk in the Early Career Spotlight Track at IJCAI'16 in New York.
In Jul'16 our STARAI'16 paper titled Contextual Symmetries in Probabilistic Graphical Models received the best paper award.
In Jun'16 I was elected as a councilor to AAAI Executive Council for a three year term.
In Apr'16 I was awarded a Young Faculty Research Fellowship under the Visvesvaraya PhD scheme for Electronics & IT by Govt. of India.
In Apr'16 I was interviewed by ML India. The transcript of the interview.
In Jan'15 at AAAI'15, I was awarded the AAAI Senior status, a distinction in the field of artificial intelligence.
In Jan'15 I was awarded a Teaching Excellence Award for my Spring 2014's AI course.
In Sep'14 I appeared on NDTV Profit to defend Artificial Intelligence at a debate show titled, The Contrarian.
In Nov'13 at HCOMP, our paper titled Crowdsourcing Multi-Label Classification for Taxonomy Creation received the best paper award.
In Oct'13 I joined as a faculty member at IIT Delhi after a six year research faculty stint at University of Washington, Seattle.
In Jul'12 Andrey Kolobov and I released a monograph titled Planning with Markov Decision Processes: An AI Perspective.
In Sep'08 I was awarded an honorable mention for the 2008 ICAPS best disseration award.
In Oct'07 I joined University of Washington as a Research Assistant Professor.
In Aug'07 I completed my PhD thesis on stochastic planning with concurrent, durative actions.
Research
At present I am working on the following projects:
- Generative AI
- Reasoning with Generative AI: Can large language models (LLMs) such as ChatGPT reason well on their own? We find that on some problems they can. For example, in JEEBench (EMNLP'23) paper, we show that GPT4 performs admirably on IIT-JEE Advanced questions. However, for other complex reasoning tasks, it needs support of traditional AI solvers.
- Cost-Quality Optimization of Generative AI: The largest LLMs like GPT4 are really expensive. But, we may not always need GPT4 for every question. Especially, easy questions should be solved by smaller models, and harder questions by the larger ones. How to manage this (and other) such optimizations dynamically for a given question? We build a series of methods for such optimizations. Adaptive Consistency (EMNLP'23) dynamically decides the number of samples to take for high confidence. AutoMix (NeurIPS'24) triages between multiple LLMs using the power of POMDPs.
- Supervision with Generative AI: Can LLMs help for tasks when a fair amount of supervised data is available? Our first results show that specialized models trained on supervised data perform better than LLMs out of the box. Careful combinations of specialized models with LLMs can push the state of the art further.
- Intelligent Information Systems & Natural Language Processing
- Open Information Extraction: We hope to overcome the "knowledge-acquisition bottleneck" by automatically extracting information from natural language text in a domain-independent manner. We work on improving the quality of Open IE extractors by pushing their precision and recall. We recently released the code for IMoJIE (ACL'20) and Open IE 6 (EMNLP'20), neural Open IE extractors, with state of the art results. Very recently, we obtained further improvements using Gen2OIE (ACL'22). Our previous Open IE extractor (Open IE 5) is publicly released with nearly 10,000 downloads. Other progress on this work includes a better handling of compound noun expressions, numerical facts and lists of facts in a sentence. We also release an evaluation framework and dataset for better evaluation of Open IE systems (paper). I wrote short survey on the vast literature on Open IE.
- Multilingual Information Extraction: How do we train IE systems that can be run on multiple languages, especially in low resource settings? We release a new dataset called DiS-ReX (ACL'22), for training and testing distantly supervised relation extraction in four languages. We create PARE (ACL'22) -- a simple baseline model that achieves state of the art results on DiS-ReX and also monolingual distant supervision datasets. We extend this to HYDRE by incorporating modern day LLMs and show value of PARE-LLM collaboration or IE in low resource languages. We also develop a multilingual Open IE (ACL'22) system that can train a neural Open IE module in any language without any language-specific training data.
- Inference over Knowledge-Bases: Knowledge-bases are always incomplete! We develop novel inference algorithms for the task of knowledge-base completion (KBC). Our type-sensitive model adds unsupervised typing to tensor factorization to obtain strong results on several datasets. We also release the code that implements these and many other models. We propose TimePlex, a KBC model for Temporal Knowledge Bases, and then design NeuSTIP (EMNLP'23) -- a neuro-symbolic model for the task. It also designs new evaluation protocols for this important task. More recently, we propose the problem of answering regular expression queries over incomplete KBs in our AKBC'21 paper. We have also extended KBC models to multilingual KBs and have released a new model called AlignKGC. Collaboration with IIT Bombay.
- Question Answering over Knowledge-Bases: We build QA systems that convert a question to a KB query. Our focus is on building robust systems, where the system does not give incorrect answers for questions that are unanswerable on the current KB. Our other focus is to rapidly retarget a QA system on a new KB, with limited training data. Collaboration with TCS Research.
- Task-oriented Dialog Systems: We study end-to-end trainable dialog systems, which are targeted towards a certain goal, such as answering customer questions. Often, these require interaction with knowledge sources, such as a knowledge-base. In BossNet (code), we show an effective way to disentangle language and knowledge when training such systems end-to-end. CDNet improves upon by adding a constrained KB-distillation layer, leading to a much better identification of appropriate entities for an utterance. An extension of RL enables training dialog systems even in absence of KB-query annotation. In EMNLP'21 paper we release a novel task-oriented dialog dataset and in EMNLP'22 paper, a system for the troubleshooting scenario, in which the agent has to ground the dialog in a given flowchart without additional annotation. Collaboration with IBM Research.
- Applications of AI
- NLP for Material Science: Most of the world's knowledge of material compositions and their properties lie in the research papers from the field. In this project, the goal is to create the world's largest knowledge-base of materials through extraction for the scientific papers. As a start, we release MatSciBERT, a language model for material science. The code and the model has been released on the HuggingFace MatSciBERT page. Further, we release DiSCoMat (ACL'23), an information extractor that extracts material compositions from tables in scientific articles. We are currently building MatSKRAFT, the world's largest repository of materials and properties. We are also gearing up to release LLaMat, a 7B parameter foundation model for material science.
- Machine Learning for Medical Imaging: An IITD-AIIMS partnership studies computer vision over histopathological images of duodenal biopsies for Celiac disease prediction. The goal is to build a medically-explainable AI system that collaborates with physicians for best predictions. In our CVPR'23 paper we develop a detection and counting approach with state of the art results in several pathology datasets, including the one for duodenal biopsies obtained at 20x zoom. Further, we develop techniques for cell segmentation with only box level supervision in our Nature Scientific Reports'25 paper. Collaboration with AIIMS, New Delhi.
- Neuro-Symbolic Machine Learning
- Using Constraints in ML: Neural models have become the model of choice for almost all machine learning applications, such as NLP, computer vision, and speech. However, previous generation (symbolic) models, based on logic or probabilistic representations can combine with neural models to achieve further progress. In this research, we explore the value(s) that symbolic constraints offer in a neural setting. In our NeurIPS'19 paper, we demonstrate how (and why) to use symbolic constraints while training a neural model for several NLP applications. Our ICLR'21 paper and ICLR'22 paper train neural models for constrained satisfaction problems like Sudoku. Our NeurIPS'22 paper trains a neural model that automatically converts the input into an ILP with constraints, effectively separating perception (neural) and reasoning (symbolic).
- Neural Models for Probabilistic Planning: Neural models for reinforcement learning problems have achieve tremendous recent success. In this project, we study whether they can also be helpful for (Relational) Markov Decision Processes (MDPs) that are expressed in a declarative logic-based representation such as RDDL. We have written a series of papers on this topic and are excited at reviving research thread of Relational MDPs using modern neural models. In our first paper (NeurIPS'18), we show that neural models trained on a few instances of a domain can be effectively transferred to a new instance of the same domain of the same size. We extend this to transfer across problem sizes in a restricted setting and in a full blown RMDP setting (ICML'20), releasing our software SymNet. We recently release SymNet 2.0, and SymNet 3.0, extensions to SymNet with much improved results (paper (UAI'22), and paper (UAI'23)).
Hobbies
In my personal time, I can be found listening to, playing, or singing hindustani classical music. I have got the fortune of accompanying several famed vocalists on harmonium, including Pt. Vidyadhar Vyas, Vidushi Sunanda Patnaik, Us. Mashkoor Ali Khan, Smt. Bharathi Prathap, and my dear wife, Shashwati Mandal. In my previous life, I performed with a Seattle light Indian music band called Pratidhwani (my last show was Kashish in December 2012). Even before that, I was involved with Seattle's local cricket tournament where I tried my fingers at off-spinning. World cinema and cooking were my other favorite pastimes.