- for Rangers & ecologists
- IJCAI 2024 CHAIRLet experts intervene on concepts mid-search
- CV4Animals @ CVPR 2026 SEEK-CBMModern backbones, 497% better retrieval
- headed To be deployed for the elephant work
Vaibhav Balloli
AI Researcher · Computer Science Ph.D. Candidate, University of Michigan
👋 Hi, I'm an applied AI researcher and a Computer Science Ph.D. candidate at the University of Michigan. I design and deploy interactive human-AI systems across healthcare, wildlife conservation, and road safety. My research studies how to make LLM agents reliable across long conversations, shifting user intent, and real-world deployment constraints through post-training, contextual evaluation, instruction following, and test-time monitoring.
Previously, I have worked at Microsoft Research: once as a summer intern working on improving interactivity of LLM agents and then as a Research Fellow for two years working on scaling HAMS, Vasudha, and improving multilingual capabilities of LLMs at VeLLM.
Systems in the field #
What got built, where it ended up, and how far along it is.
-
Deployed 2021 – 2022
HAMS
Smartphone-based driver licence testing deployed across India, serving 440,000+ candidates.
Microsoft Research India
-
In progress 2025 – 2026
PATHFinder
A four-stage agent that builds individualised prenatal care plans aligned to ACOG's PATH guidelines, with clinician oversight.
University of Michigan
-
In progress 2023 – 2026
CHAIR & SEEK-CBM
Concept-bottleneck retrieval that lets conservation experts steer elephant re-identification mid-search.
University of Michigan · MIT
News #
-
One paper accepted to KDD 2026 in the Datasets and Benchmarks track.
-
One paper accepted to CV4Animals @ CVPR 2026, in collaboration with BeeryLab at MIT.
-
One paper accepted to FAccT 2026.
-
One paper accepted to Interactive Health 2026.
-
Started my internship at AI Frontiers, Microsoft Research.
Research trails #
Seven papers, three trails. Each starts with the people it is for, runs through the papers, and ends where the work is headed. Pick a trail to filter the papers below.
- for Birthing people, clinicians & health workers
- FAccT 2026 Where is this coming from?Ask them first: four focus groups, three stakeholder groups
- Interactive Health 2026 PATHFinderBuild to that spec: a care-planning agent
- KDD 2026 RELIANCEThen measure the ecosystem it lives in
- headed TBD in Michigan
Research #
I build interactive AI systems that people can steer: agents that clarify intent and recover from errors, evaluations grounded in domain expertise, and models that adapt at run time. My work connects large language models, reinforcement learning, and human-AI interaction in healthcare, wildlife conservation, and road safety:
- Agent post-training: helping agents clarify intent, recover, abstain, and delegate across extended interactions.
- Contextual evaluation: developing benchmarks and judgments with domain experts.
- Instruction following and monitoring: studying agent behavior and tracking systems at test time.
Want to see the research in motion? See the interactive demos →
RELIANCE: Curating and Evaluating Reproductive Health Information on Social Media
Multi-modal agents curate TikTok reproductive health videos, and clinicians and LLMs judge them for accuracy and harms.
Devise data curation multi-modal agents to collect and curate reproductive health information from TikTok and work with clinicians to evaluate the accuracy of the content and how LLMs perform in evaluating the content for accuracy and harms.
Deploying trustworthy health AI systems
@inproceedings{balloli2026reliance,
title = {{RELIANCE}: Curating and Evaluating Reproductive Health Information on Social Media},
author = {Balloli, Vaibhav and Ellis, Laura Peyton and Mishra, Vishala and Chi, Alice M and Friedman Peahl, Alex and Bondi-Kelly, Elizabeth},
booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
year = {2026},
doi = {10.1145/3770855.3817447},
url = {https://doi.org/10.1145/3770855.3817447},
eprint = {2606.18285},
archivePrefix = {arXiv},
primaryClass = {cs.IR}
}
PATHFinder Agent for Tailored Prenatal Care
LLM agents that talk with pregnant people to plan personalised prenatal care and resources, and to assist clinicians.
Interactive LLM agents that converse with pregnant people to provide personalized prenatal care information, planning, and resources to patients, and to assist clinicians.
Deploying trustworthy health AI systems
@inproceedings{balloli2026pathfinder,
title = {PATHFinder Agent for Tailored Prenatal Care},
author = {Balloli, Vaibhav and Samuel, Carissa and Abdelnabi, Samia and Peahl, Alex and Bondi-Kelly, Elizabeth},
booktitle = {Proceedings of the 2026 ACM Interactive Health Conference},
year = {2026},
doi = {10.1145/3786579.3804996},
url = {https://doi.org/10.1145/3786579.3804996},
eprint = {2607.24768},
archivePrefix = {arXiv},
primaryClass = {cs.CY}
}
"Where is this coming from?" Uncovering Trustworthiness Ideals in AI-powered Peripartum Information Seeking
Focus groups on the role of AI information tools in peripartum care and what stakeholders expect for trust and governance.
Conduct focus groups to understand the role of AI-enabled information provision in the current health ecosystem and the trustworthiness and governance expectations of stakeholders.
Deploying trustworthy health AI systems
@inproceedings{balloli2026sociotechnical,
title = {``Where is this coming from?'' Uncovering Trustworthiness Ideals in {AI}-powered Peripartum Information Seeking},
author = {Balloli, Vaibhav and Erickson, Julia and Li, Xinyi and MacMurray van Liemt, Erin and Friedman Peahl, Alex and Bondi-Kelly, Elizabeth},
booktitle = {Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency},
year = {2026},
doi = {10.1145/3805689.3812277},
url = {https://doi.org/10.1145/3805689.3812277},
eprint = {2606.10158},
archivePrefix = {arXiv},
primaryClass = {cs.HC}
}
SEEK-CBM: Editable and Interpretable Retrieval for Elephant Re-Identification
Upgrades CHAIR to modern vision architectures and self-supervised training, improving performance by 497%.
SEEK-CBM upgrades CHAIR to modern vision architectures and self-supervised techniques to achieve a 497% improvement in performance.
Helping experts steer retrieval
@inproceedings{seekcbm2026salaun,
title = {{SEEK}-{CBM}: Editable and Interpretable Retrieval for Elephant Re-Identification},
author = {Sala{\"u}n, Antoine and Balloli, Vaibhav and Haucke, Timm and Bondi-Kelly, Elizabeth and Beery, Sara},
booktitle = {6th Workshop on CV4Animals: Computer Vision for Animal Behavior Tracking and Modeling, in conjunction with CVPR 2026},
year = {2026}
}
Bridging the Gap: Dynamic Learning Strategies for Improving Multilingual Performance in LLMs
Picks the prompt, LLM, and embedding model per query at run time to improve multilingual QA, with no extra training.
A dynamic strategy that selects, per query at run-time, the optimal prompt, LLM, and embedding model to maximize multilingual QA performance, improving non-Latin and low-resource language coverage without any extra training.
Selecting best system configuration
@inproceedings{kumar2025leap,
title = {Bridging the Gap: Dynamic Learning Strategies for Improving Multilingual Performance in LLMs},
author = {Kumar, Somnath and Balloli, Vaibhav and Ranjit, Mercy and Ganu, Tanuja and Ahuja, Kabir and Sitaraman, Sunayna and Bali, Kalika},
booktitle = {Proceedings of the International Conference on Computational Linguistics (COLING)},
year = {2025},
note = {Oral}
}
Are They the Same Picture? Adapting Concept Bottleneck Models for Human-AI Collaboration in Image Retrieval
Concept bottleneck models that let experts intervene on concepts mid-retrieval, for wildlife re-identification platforms like ElephantBook.
CHAIR adapts Concept Bottleneck Models so experts can intervene on intermediate concepts during image retrieval, a setting that matters for wildlife re-identification platforms like ElephantBook. Flexible levels of intervention let domain experts spend effort where it's most useful, achieving genuine human-AI complementarity.
Helping experts steer retrieval
@inproceedings{balloli2024chair,
title = {Are They the Same Picture? Adapting Concept Bottleneck Models for Human-AI Collaboration in Image Retrieval},
author = {Balloli, Vaibhav and Beery, Sara and Bondi-Kelly, Elizabeth},
booktitle = {Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, {IJCAI-24}},
editor = {Kate Larson},
publisher = {International Joint Conferences on Artificial Intelligence Organization},
pages = {7824--7832},
year = {2024},
month = {8},
note = {Human-Centred AI},
doi = {10.24963/ijcai.2024/866},
url = {https://doi.org/10.24963/ijcai.2024/866}
}
Chanakya: Learning Runtime Decisions for Adaptive Real-Time Perception
Learns run-time choices of resolution, model, and compute from scene content and system contention to balance accuracy and latency.
A learned approximate-execution framework for streaming perception. Chanakya jointly considers scene content and system contention to pick run-time decisions (resolution, model, compute) that balance accuracy and latency, beating static and dynamic baselines on both server GPUs and edge devices.
Selecting best system configuration
@inproceedings{ghosh2023chanakya,
title = {Chanakya: Learning Runtime Decisions for Adaptive Real-Time Perception},
author = {Ghosh, Anurag and Balloli, Vaibhav and Nambi, Akshay and Singh, Aditya and Ganu, Tanuja},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
year = {2023}
}
Interactive research demos #
Animated walkthroughs of how the systems below actually work, drawn from the papers themselves.
SEEK-CBM #
Editable and interpretable retrieval for elephant re-identification.
Read the paper →PATHFinder #
Interactive agents for tailored prenatal care.
View the project →Things I keep going back to #
A short, evolving shelf of books and talks I find myself returning to.
📚 Reading
-
A problem-solving-as-survival sci-fi novel: first-principles reasoning under extreme constraints.
-
On cultivating sustained, distraction-free focus as a competitive advantage for knowledge work.
-
The foundational paper on designing systems where humans and automated agents collaborate fluidly.
-
Short essay arguing that general methods leveraging computation tend to win out over hand-crafted knowledge.
🎬 Talks & videos
-
Prof. Randy Pausch's wise, funny lecture on living with intention.
-
On doing hard things for decades and betting on outsized, asymmetric outcomes.