Kartik Mehta is a Senior Applied Scientist at Amazon AGI, specializing in Large Language Models (LLMs), Natural Language Processing (NLP), and Generative AI. He leads post-training and evaluation for the Amazon Nova family of frontier models — featured in Fortune and TechCrunch. An IIT Delhi alumnus (B.Tech & M.Tech in Machine Learning), he has over a decade of experience in AI research and applied ML.
His work spans LLM reasoning, data synthesis, and alignment, including DeCRIM (EMNLP 2024), FLAMES (EMNLP 2025), and DiCoRe (EMNLP 2025). He holds a U.S. patent on entity-extraction systems deployed at Amazon scale and has authored 10+ papers in venues such as EMNLP and NAACL. Kartik also serves as a reviewer for top AI conferences and a judge for the Alexa Prize (SocialBot and TaskBot) and Amazon Nova AI Challenge (Trusted Software Agents track).
Previously, he led research in e-commerce attribute extraction (NAACL 2022, NAACL 2021; ACL Workshop 2021) and product question answering (EMNLP 2019; WWW 2019), developing systems used globally by millions of Amazon customers.
Amazon Nova (Amazon Technical Reports), FLAMES (EMNLP 2025), DiCoRe (EMNLP 2025), DeCRIM (EMNLP 2024)
NER_MQMRC (NAACL 2022, US patent granted), LATEX-Numeric (NAACL 2021), SANTA (ACL Workshop 2021)
Hard Negative (EMNLP 2019), ProductQnA (WWW Workshop 2019)
EMNLP, 2025
How do synthetic data choices truly shape LLM reasoning? Our EMNLP 2025 paper, FLAMES, introduces the first unified framework to systematically compare and optimize synthetic data pipelines—revealing what really drives performance in math reasoning models.
EMNLP, 2025
How can LLMs reason under strict structural constraints? Our EMNLP 2025 paper, DiCoRe, introduces a Divergent–Convergent reasoning framework that enables zero-shot event detection with higher accuracy, stronger constraint adherence, and 15–55× lower compute cost than standard CoT methods. (Work supported as part of Amazon Research Fellowship)
EMNLP, 2024
Do LLMs truly follow complex human instructions? Our EMNLP 2024 paper, DECRIM, introduces a self-correction framework that helps models decompose, critique, and refine multi-constraint tasks—boosting open LLM performance beyond GPT-4 on real-world instructions.
Reviewer for leading AI and NLP conferences — ARR (ACL 2025, EMNLP 2025), NAACL 2025, COLM 2024, ICML 2023, ACL 2023, EMNLP 2022.
Judged finalists on LLM safety and misuse prevention, helping allocate $700K in research awards.
Amazon Nova AI Challenge →Judged university teams competing for $1M in prizes advancing conversational AI.
Alexa Prize SocialBot →Reviewed global applications for a $500K competition on task-oriented chatbots.
Alexa Prize Taskbot →Mentored a PhD student at UCLA in GenAI/LLM Information Extraction, resulting in an EMNLP 2025 paper.
Amazon Research Awards →Recognized among the top 10% of IEEE's 400,000+ members worldwide for professional excellence and impact.
IEEE Senior Member →