Unraveling the cognitive patterns of Large Language Models through module communities(Link)
PI: Dr. Pin-Yu Chen, IBM; Dr. Jianxi Gao, RPI
Arxiv 2025
· Utilize network analysis to map the association between datasets and cognitive skills within LLMs, revealing the distribution of skills across different model modules.
· Analyze the localization of skills in LLM modules and evaluate how targeted fine-tuning of modules based on specific skill distributions impacts model performance.
· Finetuned a total of 186 models based on each community of skills: (Link)
CRAFT: Training-Free Cascaded Retrieval for Tabular QA(Link)
Co-Author: Adarsh Singh | PI: Dr. Jianxi Gao, RPI; Dr. Soham Dan, Microsoft; Dr. Vivek Gupta, Arizona State University
Arxiv 2025
· Developed a multi-stage retrieval pipeline combining sparse (SPLADE), dense (Sentence-Transformer), and neural reranking (text-embedding-3-small) modules for table question answering without fine-tuning.
· Achieved state-of-the-art retrieval performance on the NQ-Tables benchmark, with robust results under query paraphrasing and significant token efficiency gains using mini-table inputs.
· Integrated retrieval with LLMs (Mistral-7B, LLaMA3-8B, Qwen2.5-7B) to deliver superior end-to-end QA performance compared to training-heavy baselines.
Forecasting Open-Weight AI Model Growth on HuggingFace(Link)
PI: Dr. Pin-Yu Chen, IBM; Dr. Jianxi Gao, RPI
SCI-FM @ ICLR 2025
· Study the trajectory of growth of a number of fine-tuned models after release.
· Compare the impact of different companies on the AI community through their released models.
Boosting Reinforcement Learning for Network Analysis with Data Augmentation Strategies
PI: Dr. Pin-Yu Chen, IBM; Dr. Sholom Havlin, Bar-Ilan University; Dr. Jianxi Gao, RPI
· Created a robust RL and GNN framework as a medium for finding critical nodes in complex networks.
· Analyzed synthetic and real-world graphs on different critical node problems using targeted attacks.
Exploring the Robustness of Language Models for Tabular QA via Attention Analysis (Link)
Co-Author: Sixue Xing | PI: Dr. Soham Dan, IBM; Dr. Jianxi Gao, RPI
TMLR 2025
· Evaluated large language models across structural and value-based perturbations on tabular QA benchmarks (WTQ, TAT-QA, SCITAB) to assess robustness and domain bias
· Conducted detailed attention-entropy analysis, revealing strong correlations between structural perturbations, mid-layer attention shifts, and performance drops
· Benchmarked instruction-tuned and base variants of multiple models (Llama2, Llama3, Qwen, Mistral) under few-shot settings, highlighting scale and tuning effects on structured reasoning
· Alternate name: On the Robustness of Large Language Models for Tabular Question Answering
Convolutional Neural Network-based EEG Emotion Classification with the Forward Selection Wrapper technique for Channel Selection (Link)
Advisor: Dr. Jon G. Sigurjonsson, University of Iceland
JSTOR 2020
· Used forward selection wrapper technique for channel filtration.
· Used Keras to create a Convolutional Neural Network (CNN) that classifies emotion using the DEAP dataset.
Web Development Intern, Temple Sinai, 208 Summit Avenue, Summit, NJ [July 2019- Sept 2019]
- Build and deploy a website using React as a tool in a suitable web hosting platform
- Explore ideas with the client to come forward with an appealing website
Web App and App Developer, Hazelnut [June 2019 – Feb 2020]
-Front-end engineering team for implementing UI design of the website and the mobile app
-Assist Hazelnut core developers with application beta-testing and application deployment
Student Technician, IT Department, Caldwell University, Caldwell, New Jersey [Aug 2018 – May 2021]
- Provided timely resolutions to your technology inquiries
- Fix technical issues in faculty, staff, or student’s computer devices
Peer Tutor, Academic Success Center, Caldwell University, Caldwell, New Jersey [Sept 2017- Dec 2020]
- Tutored Mathematics and Computer Science courses to students in need
- Recorded and analyzed data of all the tutors to create a report for the Supervisor