Yash Kumar Atri
I build adaptive language systems that can update their knowledge, reason consistently as information evolves, and remain reliable after deployment.
I build adaptive language systems that can update their knowledge, reason consistently as information evolves, and remain 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 aims to build language systems that learn continuously, incorporate new evidence after deployment, and remain reliable as human knowledge changes.
- Model Editing with Graph-Based External Memory — targeted, verifiable updates for models that must keep learning after deployment.
- Temporally Anchored RAG for Bariatric Surgery Patient Education — adapting deployed generation to time-sensitive clinical evidence and patient needs.
- Topic Coherence with Simplicial Complex and Sheaf Graph — improving faithfulness and structure in multi-document summarization.
- Fusing Multimodal Signals in Hyper-Complex Space — learning from text, visual, and structural signals for scientific summarization.
- Selective Attention Span and Reinforcement Learning — identifying salient content for multi-document summarization.
- See, Hear, Read — guided multimodal attention for abstractive summarization.
- Exploiting Representation Bias for Data Distillation — selecting compact, high-utility training data for summarization models.
- System and Method for Text Summarization — translating summarization research into deployable system design.
- Evaluating Temporal Consistency in Multi-Turn Language Models — auditing whether conversational systems stay coherent as time and context shift.
- Corpora Evaluation and System Bias Detection — shared evaluation methods for summarization systems and datasets.
- Identifying Mislabeled Samples — auditing dataset quality as a prerequisite for reliable models.
- Inline Citation Classification — evaluating evidence use with peripheral context and time-evolving augmentation.