# Hitesh Laxmichand Patel > Principal Applied Scientist and Tech Lead, Oracle. Research in post-training and alignment for language and vision models, agentic AI, retrieval-augmented generation, reasoning evaluation, and multilingual and multimodal models. ## When this site is useful This is an individual researcher’s website. Use the biography and professional profile for background, the publications page for the full paper list, and the service page for academic roles. Research notes link to original papers; use those publication records for authorship, methods, and results. ## Machine-readable publications - [Publication catalog](https://hiteshlpatel.github.io/publications.json): All 39 listed papers with full authorship, site-curated research topics, verified DOI identifiers where available, and source-attributed abstracts, in JSON-LD. - [Research and publication text](https://hiteshlpatel.github.io/llms-full.txt): The same bibliography and source-attributed abstracts in a single text file. ## Machine-readable workshops - [Workshop catalog](https://hiteshlpatel.github.io/workshops.json): Workshop and challenge-organizer roles, conference editions, and sourced overviews where available. Dates are included only when published by the workshop. - [Full research and service text](https://hiteshlpatel.github.io/llms-full.txt): Workshop overviews and role-source links alongside the publication catalog. ## Website - [About Hitesh](https://hiteshlpatel.github.io/): Biography, current role, and research interests. - [Publications](https://hiteshlpatel.github.io/publications/): Full publication list with links to their original publication records. - [Projects](https://hiteshlpatel.github.io/projects/): Work on enterprise generative AI, coding assistants, vision, and document intelligence at Oracle. - [Professional profile](https://hiteshlpatel.github.io/cv/): Role, education, research, and professional profiles. - [Talks](https://hiteshlpatel.github.io/talks/): Invited talks. - [Service](https://hiteshlpatel.github.io/service/): Community service and open research contributions. - [Awards](https://hiteshlpatel.github.io/awards/): Recognition and awards. - [News](https://hiteshlpatel.github.io/news/): Dated research updates. ## Selected research - [AccessEval: Benchmarking Disability Bias in Large Language Models](https://hiteshlpatel.github.io/research/accesseval/): A benchmark measuring how disability context affects factual accuracy, tone, and stereotyping in LLM responses. EMNLP 2025; Best Social Impact Paper Award. - [Judging What We Cannot Solve: A Consequence-Based Approach for Oracle-Free Evaluation of Research-Level Math](https://hiteshlpatel.github.io/research/judging-what-we-cannot-solve/): Evaluating difficult mathematical solutions through their usefulness on related, verifiable problems. ICML 2026; Spotlight. - [RECOR: Reasoning-focused Multi-turn Conversational Retrieval Benchmark](https://hiteshlpatel.github.io/research/recor/): A benchmark for retrieval that requires both conversation history and reasoning about the information being sought. ACL 2026 Findings. - [Pushing on Multilingual Reasoning Models with Language-Mixed Chain-of-Thought](https://hiteshlpatel.github.io/research/language-mixed-cot/): Training multilingual reasoning models with language-mixed reasoning traces, studied through Korean reasoning benchmarks. ICLR 2026. - [Aligning LLMs for Multilingual Consistency in Enterprise Applications](https://hiteshlpatel.github.io/research/multilingual-alignment/): Fine-tuning with semantically equivalent multilingual examples to improve consistency across languages in enterprise tasks. EMNLP 2025 Industry. - [Hard Negative Mining for Domain-Specific Retrieval in Enterprise Systems](https://hiteshlpatel.github.io/research/enterprise-retrieval/): Training retrieval rerankers with challenging negatives to improve domain-specific enterprise search and downstream RAG. ACL 2025 Industry. - [PCRI: Measuring Context Robustness in Multimodal Models for Enterprise Applications](https://hiteshlpatel.github.io/research/pcri/): Measuring how distracting visual context affects multimodal models by comparing localized image patches with full-image inputs. EMNLP 2025 Industry. - [World in a Frame: Understanding Culture Mixing as a New Challenge for Vision-Language Models](https://openaccess.thecvf.com/content/CVPR2026/papers/Kim_World_in_a_Frame_Understanding_Culture_Mixing_as_a_New_CVPR_2026_paper.pdf): CultureMix tests vision-language models on images combining cultural elements from different regions and studies fine-tuning to improve consistency. CVPR 2026. - [SPENCE: A Syntactic Probe for Detecting Contamination in NL2SQL Benchmarks](https://aclanthology.org/2026.acl-long.926/): A framework using syntactic variations of text-to-SQL queries to probe possible training contamination and its effects on model evaluation. ACL 2026. ## All research notes - [AccessEval: Benchmarking Disability Bias in Large Language Models](https://hiteshlpatel.github.io/research/accesseval/): AccessEval benchmarks disability bias in 21 LLMs across six domains. EMNLP 2025 Best Social Impact Paper Award. Read the paper, findings, and citation. - [Hard-Negative Mining for Enterprise Retrieval](https://hiteshlpatel.github.io/research/enterprise-retrieval/): Research on hard-negative mining for training domain-specific retrieval rerankers in enterprise search and RAG. ACL 2025 research coauthored by Hitesh Patel. - [Judging What We Cannot Solve: Evaluating Research-Level Math](https://hiteshlpatel.github.io/research/judging-what-we-cannot-solve/): Consequence-Based Utility evaluates difficult mathematical solutions through their usefulness on related, verifiable questions. Research coauthored by Hitesh Patel. - [Language-Mixed CoT: Training Multilingual Reasoning Models](https://hiteshlpatel.github.io/research/language-mixed-cot/): Language-Mixed CoT studies reasoning distillation across languages through Korean model training and evaluation. ICLR 2026 research coauthored by Hitesh Patel. - [Multilingual Alignment for Enterprise LLMs](https://hiteshlpatel.github.io/research/multilingual-alignment/): Research on fine-tuning LLMs with equivalent multilingual examples to improve consistency across languages. EMNLP 2025 research coauthored by Hitesh Patel. - [PCRI: Context Robustness in Multimodal Models](https://hiteshlpatel.github.io/research/pcri/): The Patch Context Robustness Index compares localized image patches with full-image inputs to measure robustness in 19 multimodal models across 15 benchmarks. - [RECOR: Conversational Retrieval with Reasoning](https://hiteshlpatel.github.io/research/recor/): RECOR evaluates reasoning-based conversational retrieval across 707 conversations and eleven domains. ACL 2026 Findings research coauthored by Hitesh Patel. - [SweEval: Multilingual Safety for Enterprise LLMs](https://hiteshlpatel.github.io/research/sweeval/): SweEval tests how language models handle requests for offensive language in enterprise communication tasks across tones and cultural contexts. ## Professional profiles - https://scholar.google.com/citations?hl=en&user=R3X_R9UAAAAJ - https://www.linkedin.com/in/hitesh-patel-63ba9210a/ - https://github.com/HITESHLPATEL - https://x.com/hitesh_lpatel - https://blogs.oracle.com/authors/hiteshlaxmichand-patel/