Hitesh Laxmichand Patel
I’m a Principal Applied Scientist and Tech Lead at Oracle. My research centers on developing language and vision models with stronger reasoning and more reliable behavior across languages and cultures. This includes post-training for multilingual and multimodal models, and agents that retrieve knowledge, use tools, and collaborate over long horizons.
A central focus of my research is closing the gap between model capability and reliable behavior. My work on evaluation examines the limits of reasoning, knowledge use, and agent behavior, informing how models are trained and aligned. Safety and fairness are central to this agenda, including how well these systems serve people across languages and cultural contexts.
Our work on AccessEval received the EMNLP 2025 Best Social Impact Paper Award, and Judging What We Cannot Solve was selected as an ICML 2026 Spotlight.
I received my M.S. from New York University, advised by Prof. Ralph Grishman. I contribute to open research through Cohere For AI’s Aya initiative, SEACrowd, and CommonLID, a collaboration involving Common Crawl, EleutherAI, and MLCommons. I also mentor students, give invited talks and tutorials, and organize workshops and research challenges.
Selected papers
- AccessEval — A benchmark for how disability context affects factual accuracy, tone, and stereotyping in LLM responses.
- Judging What We Cannot Solve — Evaluating research-level mathematical solutions through their usefulness on related questions with verifiable answers.
- Language-Mixed CoT — Studying language-mixed reasoning traces for multilingual reasoning distillation, using Korean as a case study.
- World in a Frame — Studying how mixed cultural context affects vision-language models and how fine-tuning improves consistency.
- SPENCE — Probing text-to-SQL benchmarks for signs of training contamination through controlled changes to query syntax.