Nobin Sarwar

Research Fellow @ SPAR · Graduate Research Assistant @ UMBC · 1x Research Intern @ LLNL

Nobin Sarwar

I am a 3rd-year Computer Science Ph.D. student in the Language Understanding Lab at the University of Maryland, Baltimore County, advised by Prof. Francis Ferraro.

My research focuses on post-training for LLMs and multimodal models, especially agentic and multi-agent reasoning, controllable adaptation, and trustworthy evaluation. Current work includes:

  • Reasoning and verification: agentic retrieve-verify workflows for scientific claims, human-AI scientific ideation (Beyond Better Ideas), and collusion in multi-agent systems

  • Evaluation and AI safety: multicultural benchmarks (Pluralis, CulturalRiddles), jailbreak robustness, and agentic dataset documentation (CroissantMiner, NeurIPS 2026; Croissant Baker, NeurIPS 2026)

  • Private and controllable adaptation: federated fine-tuning with differential privacy (FedMentor, FedMentalCare) and multimodal unlearning (Survey, ACL 2026)

Research Scientist Intern @ LLNL (Summer 2026) — AI safety, protecting confidential data when using frontier models, scientific reasoning, and adversarial evaluation; featured in Meet LLNL Interns.

Previously, I earned my MS in Computer Science from the University of Texas Rio Grande Valley, where I worked on differentially private Federated Learning for biometrics, contributing to projects that received NSF funding.

Open to collaborations on agentic reasoning, post-training, and LLM evaluation — and to mentoring undergrad and grad students new to research. Email me anytime.

Research News (View All)

Oct 24, 2026 🤝 Our paper Beyond Better Ideas, exploring the impact of human-AI collaboration on scientific ideation, is now available on arXiv.
Oct 21, 2026 📰 CulturalRiddles is now on arXiv, with news coverage in Fast Company and the Cohere blog (posts: X; LinkedIn).
Oct 07, 2026 🥐 Our paper CroissantMiner is now available on arXiv.
Sep 25, 2026 🎉 CroissantMiner and Croissant Baker got accepted as posters at NeurIPS 2026.
Sep 03, 2026 🌐 Co-authored the Last Translation Benchmark, a community-built dataset designed to break translation models with reliable evaluation rules. Check it out here.

Publications Spotlight

Full publication list on Google Scholar →

  1. Preprint’26
    beyond_better_ideas.png
    Beyond Better Ideas: How Human-AI Collaboration Shapes Scientific Ideas and Human Perceptions
    Qian Chen, Saatvik Pradhan, Janice Ahn, Berk Atil, Angkai Li, Shijie Lu, Nobin Sarwar, and 3 more authors
    arXiv Bib PDF
  2. Preprint’26
    pluralis.png
    Pluralis v0.1: Towards a Multicultural, Multimodal, Multilingual Benchmark for AI Risk and Reliability
    Alicia Parrish, Rajat Shinde, ..., Nobin Sarwar, ..., and Lora Aroyo
    arXiv preprint arXiv:2607.06196
    arXiv Bib PDF
  3. NeurIPS’26
    CroissantMiner.png
    CroissantMiner: Automated Extraction and Validation of Croissant Metadata for ML Datasets
    Berke Arda, Mubashara Akhtar, Ahmetcan Yavuz, Paul Gerry, Sebastian Lobentanzer, Nobin Sarwar, Joan Giner-Miguelez, and 3 more authors
    In Advances in Neural Information Processing Systems (Evaluations and Datasets Track)
    arXiv Bib PDF Code Website
  4. NeurIPS’26
    croissant_baker.png
    Croissant Baker: Metadata Generation for Discoverable, Governable, and Reusable ML Datasets
    Rafi Al Attrach, Rajna Fani, Sebastian Lobentanzer, Joan Giner-Miguelez, Debanshu Das, Varuni H. K., Nobin Sarwar, and 13 more authors
    In Advances in Neural Information Processing Systems (Evaluations and Datasets Track)
    arXiv Bib PDF Code
  5. ACL’26
    multimodal_unlearning_survey.jpg
    Multimodal Unlearning Across Vision, Language, Video, and Audio: Survey of Methods, Datasets, and Benchmarks
    Nobin Sarwar, Shubhashis Roy Dipta, Zheyuan Liu, and Vaidehi Patil
    In Findings of the Association for Computational Linguistics: ACL 2026
    arXiv Bib PDF Code Website