Jyotikrishna Dass

Fine-tuning Vision Transformers (ViTs) with low-rank adapters (LoRA) promises better communication efficiency under federated setup, yet existing aggregation strategies face fundamental limitations. Independently averaging these LoRA factors is mathematically inconsistent, introducing cross-term aggregation error.

[Submitted on 26 Jun 2026] Title:GLIDE: Guided Layerwise Hybrid Attention for Efficient LLM Inference View a PDF of the paper titled GLIDE: Guided Layerwise Hybrid Attention for Efficient LLM Inference, by Vimal William and 1 other authors View PDF HTML (experimental) Abstract:As Large Language Models scale to increasingly long contexts, the memory I/O and computational overhead of the Key-Value (KV) cache during decoding emerges as the primary throughput bottleneck.

[Submitted on 10 Jun 2025] Title:FLoRIST: Singular Value Thresholding for Efficient and Accurate Federated Fine-Tuning of Large Language Models View a PDF of the paper titled FLoRIST: Singular Value Thresholding for Efficient and Accurate Federated Fine-Tuning of Large Language Models, by Hariharan Ramesh and Jyotikrishna Dass View PDF HTML (experimental) Abstract:Integrating Low-Rank Adaptation (LoRA) into federated learning offers a promising solution for parameter-efficient fine-tuning of...