cv

Postdoctoral Fellow

Mila - Quebec AI Institute

Focus: Enhancing Scientific Workflows with AI. Advisors: Prof. Chris Pal, Prof. Laurent Charlin

• Led the LitLLM project, a novel RAG pipeline for literature search, improving recall from 8.2% → 84.4% with "Deep Research" implementation and reducing hallucinations by 18-26%, enabling researchers to access relevant papers more reliably
• Created ReviewerToo, a modular, training-free AI-assisted peer review framework achieving near-human accuracy for the task of accepting/rejecting a scientific paper (81.8% vs. 83.9%), in addition to providing systematic assessments
• Developed AInstein, a self-reflective research agent that autonomously solves AI problems; independently rediscovers 20% of ICLR 2025 approaches and generates 60% novel solutions, demonstrating frontier potential for automated scientific discovery
• Built an agentic pipeline to curate Wiki-style articles for Physics-focused RAG applications covering >300k Physics papers

Visiting Researcher

ServiceNow Research

Focus: Data Augmentation & Agentic Data Analysis. Advisors: Dzmitry Bahdanau, Issam Laradji

• Proposed PromptMix, a 2-shot data augmentation strategy outpeforming 5-shot and 100-shot text classification methods
• Developed MixSumm and PPSL for text summarization, matching fully supervised methods with 5% of the labeled data
• Spearheaded the creation of InsightBench, a comprehensive benchmark with 100 diverse tasks to evaluate data analytics agents, and developed AgentPoirot, a multi-step reasoning agent that autonomously discovers complex insights from data

Graduate Research

University of Waterloo

Focus: Computational Creativity & Multi-modal AI. Advisor: Prof. Olga Vechtomova

• Proposed a novel, interpretable framework to computationally measure artistic inspiration in poetry and predict aesthetic preferences of creatives, outperforming a 450-shot LLaMA classifier by 18 points on the curated EvocativeLines dataset
• Trained a bi-modal CVAE+LSTM architecture with an adversarial loss for LyricJam Sonic, a real-time, bi-modal system for music and lyric co-creation; also implemented a BERT filter to improve the coherence of the served lyrics at runtime
• Designed fusion mechanisms (Auto-Fusion and GAN-Fusion) to adaptively combine multi-modal data sources for improved emotion recognition and machine translation; also worked on the Multi-Modal Discussion Transformer (mDT), which integrates text, image, and graph transformer data to more effectively detect hate speech in Reddit discussions
• Proposed adversarial alignment for multi-turn dialogs [16] and studied the racial and ethnic bias in bots v/s humans

Undergraduate Research

IIT Kharagpur

Focus: Linguistics-grounded NLP. Advisor: Prof. Pawan Goyal

• Designed program synthesis models for morphological inflection in English, and the first energy-based model for Sanskrit

PhD

University of Waterloo

Computer Science

• Thesis: Harnessing Generalist LLMs for Diverse Objective and Subjective NLP Tasks

M.Math

University of Waterloo

Computer Science

• Thesis: Adaptive Fusion Techniques for Effective Multimodal Deep Learning

B.Tech (Hons)

IIT Kharagpur

Manufacturing Science and Engineering (Mechanical Engineering)

• Thesis (in Computer Science): Program Synthesis for Natural Language

Balancing indeterminacy and structure: Neural text generation for artistic inspiration

International Conference on Computational Intelligence in Music, Sound, Art and Design (Part of EvoStar)

Olga Vechtomova and Gaurav Sahu.

Insightbench: Evaluating business analytics agents through multi-step insight generation

ICLR 2025

Gaurav Sahu, Abhay Puri, Juan Rodriguez, Amirhossein Abaskohi, Mohammad Chegini, Alexandre Drouin, Perouz Taslakian, Valentina Zantedeschi, Alexandre Lacoste, David Vazquez, et al.

A guide to effectively leveraging llms for low-resource text summarization: Data augmentation and semi-supervised approaches

Findings of the Association for Computational Linguistics: NAACL 2025

Gaurav Sahu, Olga Vechtomova, and Issam H Laradji.

Litllms, llms for literature review: Are we there yet?

Transactions on Machine Learning Research

Shubham Agarwal*, Gaurav Sahu*, Abhay Puri*, Issam H Laradji, Krishnamurthy Dj Dvijotham, Jason Stanley, Laurent Charlin, and Christopher Pal.

Computational modeling of artistic inspiration: A framework for predicting aesthetic preferences in lyrical lines using linguistic and stylistic features

arXiv preprint arXiv:2410.02881

Gaurav Sahu and Olga Vechtomova.

Llm aided semi-supervision for efficient extractive dialog summarization

Findings of the Association for Computational Linguistics: EMNLP 2023

Nishant Mishra, Gaurav Sahu, Iacer Calixto, Ameen Abu-Hanna, and Issam Laradji.

Promptmix: A class boundary augmentation method for large language model distillation

Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

Gaurav Sahu, Olga Vechtomova, Dzmitry Bahdanau, and Issam Laradji.

Data augmentation for intent classification with off-the-shelf large language models

Proceedings of the 4th Workshop on NLP for Conversational AI

Gaurav Sahu, Pau Rodriguez, Issam Laradji, Parmida Atighehchian, David Vazquez, and Dzmitry Bahdanau.

Mentor & 3x World Finalist

Technovation Girls

Mentored 8-18 yrs old girls to solve societal problems using technology, with teams reaching the world finals three times.

Mentor

Student Welfare Group, IITKGP

Helped 7 undergraduates successfully navigate their initial years at IIT Kharagpur.

Best Paper Award

International Conference on Computational Creativity (ICCC)

Built an interpretable framework for modeling inspiration in creative individuals.

Mitacs Accelerate Scholarship

Mitacs

$120,000

Graduate Excellence Award

University of Waterloo

$10,000

Intl Doctoral and Masters Award

University of Waterloo

$37,500

Recipient of Prime Minister Scholarship

Government of India

For passing JEE Advanced 2014 examination (< 0.1% acceptance rate)

State-level Silver medalist

6th International Mathematics Olympiad (conducted by the Science Olympiad Foundation)