Yash Kumar Atri
I build adaptive language systems: large language models that can be edited, keep learning as information evolves, and stay reliable after deployment.
I build adaptive language systems: large language models that can be edited, keep learning as information evolves, and stay reliable after deployment.
Hi! I am a postdoctoral researcher at the University of Virginia, working with Prof. Tom Hartvigsen. Before joining UVA, I completed my PhD in Computer Science & Engineering at IIIT Delhi, advised by Prof. Vikram Goyal and Prof. Tanmoy Chakraborty. My research focuses on continual learning, model editing, and evaluation of large language models, including clinical applications where outdated answers carry real risk.
Models are trained once and used for years, while the world keeps changing. My work is about three things: repairing what a model gets wrong, helping it adapt as things change, and verifying that it still works.
Fixing systems that are already deployed: correcting what a model knows without retraining it or breaking the rest.
- Lifelong Model Editing with Graph-Based External Memory ACL Findings 2025 — curvature-aware, traceable edits that survive multi-hop reasoning.
- MedKIT NeurIPS 2026 D&B — how editing and update methods absorb new clinical facts.
- Topic Coherence with Simplicial Complex and Sheaf Graph EMNLP 2023 — structure that keeps information consistent across documents.
Continual learning: models that keep taking in new information and stay consistent as the world changes.
- Temporally Anchored RAG for Bariatric Surgery Patient Education Obesity Surgery 2026 — retrieval that tracks time-sensitive clinical guidance.
- Inline Citation Classification PAKDD 2023 — evidence use with time-evolving augmentation.
Testing whether updates actually hold up, especially in medicine.
- MedKIT NeurIPS 2026 D&B — generalization and knowledge preservation after clinical updates.
- Temporal Consistency in Multi-Turn Language Models ACL 2026 — auditing coherence as time and context shift.
- Corpora Evaluation and System Bias Detection EMNLP Findings 2020 — shared evaluation for summarization systems and datasets.
- Identifying Mislabeled Samples ASONAM 2021 — dataset quality as a prerequisite for reliable models.