Avatar

Somak Aditya

Assistant Professor

Department of CSE, IIT KGP

Biography

Latest News
  • I am looking for motivated full-time Ph.D students who are interested in pursuing their thesis in the exciting juncture of classical and statistical AI (often dubbed as neuro-symbolic methods).

I am an Assistant Professor (Grade I) at IIT Kharagpur Department of CSE. I was a Postdoctoral Researcher at Microsoft Research India advised by Dr. Monojit Choudhury. Prior to joining MSR India, I spent 1.5 years in Adobe Research as a full-time Researcher. I completed by Ph.D from CIDSE, Arizona State University in June 2018, under the supervision of Prof. Chitta Baral and Prof. Yezhou Yang.

Deployable AI systems should be able to reason with knowledge that is commonplace to humans. Thus, my research aims to enhance, evaluate, and explain different types of complex reasoning abilities of AI systems. At MSRI, I explored evaluation and enhancement of reasoning capabilities of Transformers-based language models. Apart from end-to-end reasoning, I am paralley exploring multi-hop reasoning capabilities of neural methods in both symbolic and natural language domains. During my Ph.D, I have explored enhancement of reasoning capabilities of image understanding systems. Through a combination of Deep Learning, Knowledge Representation, and Probabilistic Logical Reasoning, I demonstrated the benefits of using reasoning and knowledge in Visual Question-Answering, Captioning, image puzzle solving and visual reasoning.

News

Sep 2024 CONLL 2024: One Long Paper accepted in Main Track about Text2Afford: Probing Object Affordance Prediction abilities of Language Models solely from Text (with Sayantan Adak, Daivik Agarwal, Animesh Mukherjee).
Sep 2024 EMNLP 2024: Two Long Papers accepted in Main Track about i) Code Prompting Elicits Conditional Reasoning Abilities in Text+Code LLMs (with Haritz Puerto, Martin Tutek, Xiaodan Zhu, & Iryna Gurevych), ii) ERVQA: A Dataset to Benchmark the Readiness of Large Vision Language Models in Hospital Environments (Sourjyadip Ray, Kushal Gupta, Soumi Kundu, Dr Payal Arvind Kasat, Somak Aditya, Pawan Goyal)
June 2024 Invited for a talk at Saarland University on "The lack of reasoning ability in LLMs, the promise of Neurosymbolic NLP, and the pitfalls of "both" (Host: Dr. Michael Hahn).
June 2024 Invited for a talk at the Probability for AI-ML-NLP series, IIT Bombay (Host: Dr. Pushpak Bhattacharya).
May 2024 Invited for a talk at UKP Day (Host: Dr. Iryna Gurevych).
May 2024 TU Darmstadt: Visiting UKP Lab, TU Darmstadt for the summer (Host: Dr. Iryna Gurevych).
Mar 2024 NAACL 2024: Long Paper accepted in Main Track about Tool-Augmented LLMs and Mathematical Reasoning MATHSENSEI: A Tool-Augmented Large Language Model for Mathematical Reasoning with Debrup Das, Debopriyo Banerjee, and Ashish Kulkarni. A shorter draft previously also published in ICLR 2024 ME-FOMO Workshop.
Feb 2024 Tr$^2$AIL Lab: Tr$^2$AILer Atharva Naik received PhD offers from CMU, USC, UMass Amherst and Cornell.
Feb 2024 Tr$^2$AIL Lab: Tr$^2$AILer Debrup Das received PhD offers from UIUC and UM Amherst.
Feb 2024 ECAI 2024: Serving as an SPC member.
Feb 2024 LREC-COLING 2024: 1 Paper accepted about Jailbreaks and LLMs Tricking LLMs into Disobedience: Understanding, Analyzing, and Preventing Jailbreaks with Abinav, Atharva, Sachin and Monojit.
Jan 2024 Google Research @Week: Spent a wonderful time at Google Research India with stalwarts from the Indian AI community. Check out posts by Jeff Dean on Twitter.
Jan 2024 Preprints: Two new preprints released on numeric and conditional reasoning: 1) "STUCK IN THE QUICKSAND OF NUMERACY, FAR FROM AGI SUMMIT: EVALUATING LLMS’ MATHEMATICAL COMPETENCY THROUGH ONTOLOGY-GUIDED PERTURBATIONS" with SUTD, University of Michigan and 2) Code Prompting Elicits Conditional Reasoning Abilities in Text+Code LLMs" with UKP Lab, TU Dramstadt, Queen's University.
Dec 2023 Microsoft: Thanks Microsoft for awarding additional Azure grant of USD 20000 under the "Accelerate Foundation Models Academic Research" initiative.
Nov 2023 CoLM 2023: Invited as Area Chair in the new Conference on Langugage Models! Please submit your papers.
Dec 2023 Microsoft: Thanks Microsoft for awarding additional Azure grant of USD 20000 under the "Accelerate Foundation Models Academic Research" initiative.
Nov 2023 CoLM 2023: Invited as Area Chair in the new Conference on Langugage Models! Please submit your papers.
Nov 2023 Tr$^2$AIL Lab: Tr$^2$AILer Sachin Vashishtha won the PMRF scholarship for 2023! Congratulations Sachin!
Nov 2023 Tr$^2$AIL Lab: Tr$^2$AILer and MTech student Debrup Das's internship with Rakuten India got extended for the second time. He is working now for 8 months funded by Rakuten! Congratulations Debrup!!
Sep 2023 AACL-IJCNLP 2023: Two Long papers accepted in Main Conference: 1) SYNC: A Structurally guided Hard Negative Curricula for Efficient Neural Code Search (with Atharva Naik, Soumitra Das, Jyothi Vedurada), 2) Prover: Generating Intermediate Steps for NLI with Commonsense Knowledge Retrieval and Next-Step Prediction (with Deepanway Ghosal, Monojit Choudhury). Stay tuned for code and preprints!
Aug 2023 LREV 2023: LREV Journal on Logical NLI dataset accepted with Ishan Tarunesh and Monojit Choudhury (Updated preprint soon).
Jun 2023 EMNLP 2023: Invited as an Area Chair.
Jun 2023 AAAI 2024: Invited as a Senior PC member for the second time.
May 2023 Microsoft: Received the "Accelerate Foundation Models Academic Research" 2023 by Microsoft jointly with Prof. Animesh Mukherjee. This comes with large Azure credits and OpenAI API access.
May 2023 ACL 2023: Our work on "A Robust Information-Masking Approach for Domain Counterfactual Generation" with Prof. Soujanya Poria, Pengfei Hong, Rishabh Bhardwaj, and Navonil Majumdar. Full credit goes to my co-authors!!!
Apr 2023 Ashoka: I was invited to give a talk at Ashoka University for the AI@Ashoka Workshop. The talk slides are available here: AI Progress & Hype under the lens of Reasoning (PDF 14.0 MB size).
Mar 2023 CS60092 COURSE UPDATE: Recording of GUEST LECTURE BY Dr. Anirudh Kembhavi, Director of Computer Vision at Allen Institute for AI (AI2) is available here.
Jan 2023 Toloka AI Grant: Received the Toloka AI Annotation Grant of USD 300 for data creation in the Toloka AI platform. Thanks Toloka AI!
Jan 2023 IIT KGP FSRG: Received the FSRG Grant aproval from IIT Kharagpur on "The Role of Feedback in Vision-Language enabled Embodied Agents towards Application to Disaster Management"!
Jan 2023 USPTO: Second patent granted (filed during Adobe Research) on Marketing Knowledge Graph Creation! Congratulations to my co-author (/co-inventor) Dr. Atanu Sinha!
Dec 2022 IndoML 2022: My invited talk on "The Relevance of Representation and Reasoning in the era of (Large) Language Models" is available here.
Sep 2022 AACL-IJCNLP 2022: Two papers accepted. "Multilingual CheckList: Generation and Evaluation" in AACL Findings and "Vector Space Interpolation for Query Expansion" as a short paper. Kudos to the all my co-authors! Stay tuned for preprints.
Sep 2022 SERB DST: Startup Research Grant on "Learning from Rules and Data for Image Analytics" got Approved!
Aug 2022 New Prerpint: Generating Intermediate Steps for NLI with Next-Step Supervision. We show how intermediate steps/proof tree can help. The work is under review. Code and website to be public soon.
Jul 2022 AAAI 2023: Honored to be serving as a Senior PC member for AAAI 2023!
Apr 2022 IIT KGP: I was promoted to Assistant Professor Grade-I (effective from Nov 29, 2021).
Feb 2022 Rakuten Institute of Technology: Invited Talk on "Evolution of Representation and Reasoning in the era of Deep Learning".
Jan 2022 Google India: USD 5K Google Cloud Credits Award.
Dec 2021 MSR India SNLP Group: Invited End-of-Postdoc Talk.
Nov 2021 IIT KGP: Joined IIT KGP CSE as Assistant Professor.
Oct 2021 AAAI 2022 (Demo): LITMUS Predictor: An AI Assistant for Building Reliable, High-performing and Fair Multilingual NLP Systems accepted in AAAI 2022 Demonstrations track (joint work with the Speech and NLP group at MSR India).
Oct 2021 EMNLP 2021 (MRL Workshop): TaxiXNLI dataset released publicly!
Sep 2021 EMNLP 2021 (MRL workshop): Work on "Analyzing the Effects of Reasoning Types on Cross-Lingual Transfer Performance" accepted, with co-authors Karthikeyan K, Aalok Sathe and Monojit Choudhury.
Aug 2021 IndoML 2021: Invited to be a part of the organizing committee (specifically datathon@IndoML). See IndoML 2020 for reference.
Aug 2021 NeurIPS-2021 MathAI4ED: Invited as a PC member.
Aug 2021 Invited Lecture at IEEE SPS Deep Learning Summer School on "The Interplay between Deep Learning, Logic, and Reasoning".
Jun 2021 CIKM 2021: Workshop CfP for "Knowledge Injection in Neural Networks" is now public! Please consider submitting!!
Jun 2021 Seminar at IIT-Hyderabad CSE on "Explicit Knowledge and Reasoning in the era of Deep Learning".
May 2021 Seminar at IIT-Kharagpur CSE on "Explicit Knowledge and Reasoning in the era of Deep Learning".
May 2021 Seminar at IIT-Kharagpur Center for AI on "Explicit Knowledge and Reasoning in the era of Deep Learning".
Mar 2021 ICLR 2021 (MathAI workshop): Work on "Analyzing The Nuances of Transformers' Polynomial Simplification Abilities" accepted, with co-authors Vishesh Agarwal, and Navin Goyal.
Mar 2021 IJCAI 2021: Workshop CfP for "Is Neuro-symbolic SOTA still a myth for NLI" is now public! Please consider submitting!!
Mar 2021 IJCAI 2021: Workshop proposoal on "Is Neuro-symbolic SOTA still a myth for NLI" accepted with co-organizers Maria Chang, Swarat Chaudhuri, Monojit Choudhury, and Sebastijan Dumancic.
Mar 2021 ICLR 2021 (MathAI workshop): Invited as PC member.
Dec 2020 ACL-IJCNLP-2021: Invited as PC member.
Nov 2020 CoNLL 2020: Our pre-recorded talk on TaxiNLI is online. Hope you enjoy (modulo my lack of expressions :))!
Oct 2020 NAACL-2021: Invited as PC Member.
Sep 2020 CoNLL-2020: TaxiNLI: Taking a ride up the NLU Hill accepted at CoNLL.
Aug 2020 EACL-2021, EMNLP-2020, LANTERN@COLING-20: Invited as PC Member.
Feb 2020 Joined MSR India as a Postdoc Researcher.
Jan 2020 ACL-2020: Invited as PC Member.
Dec 2019 AAAI-20: "Uncovering Relations for Marketing Knowledge Representation" accepted in StarAI workshop full paper.

Projects

NeuroSymbolic Reasoning

We investigate how to combine neural networks, logical systems to perform complex reasoning (agnostic of modalities)

Mathematical Reasoning

We explore multi-hop reasoning capabilities of Transformer/GNN methods in purely symbolic domains.

Reasoning in NLP

We investigate how to evaluate, explain and enhance neural models under the lens of reasoning.

Vision and Reasoning

In this project, we explore how reasoning with knowledge enhance image understanding.

Contact