Research Publications

My research spans large language models, generative AI, personalization systems, and LLM alignment. Below are my published and preprint papers.

Iterative Critique-Refine Framework for Enhancing LLM Personalization

arXiv Preprint | October 2025

Developed PerFine, a training-free critique-refine framework for personalized text generation. The system enhances user-specific tone, style, and topical alignment through iterative feedback between a generator and a critic model. Tested across Yelp, Goodreads, and Amazon datasets, PerFine achieved 7-13% improvements in personalization quality over prior RAG-based methods.

LS-GAN: Human Motion Synthesis with Latent-space GANs

IEEE WACV Workshop 2025 | February 28, 2025

Generated text-to-motion sequences in latent space utilizing GANs, VAE, and CLIP (Distributed training) on HumanML3D and HumanAct12. Achieved a FID of 0.48 with GAN in the latent space with 91% in FLOPs reduction compared to latent diffusion model on HumanML3D. This work was accepted and published at the IEEE Winter Conference on Applications of Computer Vision (WACV) 2025 Workshop.

Safe to Serve: Aligning Instruction-Tuned Models for Safety and Helpfulness

arXiv Preprint | November 2024

Aligned LLaMA-2 7B toward safety using PEFT techniques (LoRA, QLoRA) on PKU-SafeRLHF benchmark with SFT, RAFT, RLHF, and DPO in Unsloth and TRL. Scored 93% safe on DPO model (compared to 40% on SFT) with Llama-Guard on I-CoNa benchmark. Additionally, DPO achieved 63.3% performance compared to SFT's 60.38% on PIKA. Implemented LLM-as-a-judge methodology to evaluate both safety and helpfulness of aligned models.

Automated Model Selection for Tabular Data

arXiv Preprint | January 2024

Structured data in the form of tabular datasets contain features that are distinct and discrete, with varying individual and relative importances to the target. Combinations of one or more features may be more predictive and meaningful than simple individual feature contributions. R's mixed effect linear models library allows users to provide such interactive feature combinations in the model design. However, given many features and possible interactions to select from, model selection becomes an exponentially difficult task. We aim to automate the model selection process for predictions on tabular datasets incorporating feature interactions while keeping computational costs small. The framework includes two distinct approaches for feature selection: a Priority-based Random Grid Search and a Greedy Search method. The Priority-based approach efficiently explores feature combinations using prior probabilities to guide the search. The Greedy method builds the solution iteratively by adding or removing features based on their impact. Experiments on synthetic datasets demonstrate the ability to effectively capture predictive feature combinations.