Research asking how models think, and how to make that thinking legible.
The Core Question
How do you run a large language model on a phone without gutting its intelligence? This paper proposes an answer.
Formal Record
“Fragile Mastery: Are Domain-Specific Trade-Offs Undermining On-Device Language Models?”
AuthorsB Jha, F Paudel
VenuearXiv preprint arXiv:2503.22698
Methodology
Proposes the Generalized Edge Model (GEM), a framework for optimizing on-device language models across resource-constrained edge devices.
The Core Question
Can a model tell you not just what it concluded, but why? This paper is about making that legible.
Formal Record
“Thinking About Thinking: SAGE-nano's Inverse Reasoning for Self-Aware Language Models”
AuthorsB Jha, F Paudel, U Puri, Z Yuting, C Donghyuk, W Junhao
VenuearXiv preprint arXiv:2507.00092
Methodology
Proposes Inverse Reasoning, a framework for improving model interpretability and reasoning transparency.
The Core Question
Medical images are hard for models to reason about efficiently. This paper adapts a vision-language model to do it with far less compute.
Formal Record
“LoRA-Enhanced PaliGemma for Efficient Visual Question Answering in Gastrointestinal Imaging”
AuthorsPK Jha, F Paudel, D Jha
VenueMediaEval 2025 (Peer Reviewed)
Methodology
Adapts Google's PaliGemma VLM using Parameter-Efficient Fine-Tuning (LoRA) for visual question answering in clinical gastrointestinal diagnostics.