RESEARCH

Working Papers

• Automated Design and Optimization of Enterprise-Scale AI Agent Systems

Co-Author: Ling Yue | PI: Dr. Jianxi Gao, RPI; Dr. Shaowu Pan, RPI; Dr. Pin-Yu Chen, IBM; Dr. Irene Ko, IBM
This project develops AGO, a framework that automatically discovers and improves enterprise-scale multi-agent workflows by searching over both workflow structures and node-level settings such as prompts, tools, memory, and decoding strategies. It builds candidate systems through a constrained supergraph and a JSON-based DSL with validation rules for structure, safety, and budget, then evaluates them on mcp-bench using metrics such as success, cost, latency, and risk. The work also compares training-free baselines with learning-based neural architecture search methods to improve robustness and success under budget limits, and validates results across benchmarks, including AssetOpsBench and MCP-Bench, with open-source LLMs such as Llama 4 and gpt-oss.

• Improving Robustness of Tabular Retrieval via Representational Stability

Co-Author: Adarsh Singh | PI: Dr. Jianxi Gao, RPI; Dr. Soham Dan, Microsoft; Dr. Vivek Gupta, Arizona State University

This project studies the instability of table retrieval and shows that performance is highly sensitive to how tables are serialized, with large recall@1 variation across datasets such as WTQ, WikiSQL, and NQ Tables. It identifies centroid-based embeddings as a strong and consistent retrieval target across several retrievers, including BGE-M3, MPNet, ReasonIR, and SPLADE. Building on that finding, the project introduces a VICReg-inspired post-hoc adapter for frozen embedding models that moves single-view embeddings toward a centroid-informed space using identity, invariance, variance, and covariance objectives, with the goal of making retrieval more accurate and stable.

Forthcoming and Under Review

• Unraveling the cognitive patterns of Large Language Models through module communities (arXiv, 2025)

PI: Dr. Pin-Yu Chen, IBM; Dr. Jianxi Gao, RPI
CogInterp @ NeurIPS 2025  and UniReps @ NeurIPS 2025 

Under Review @ Proceedings of the National Academy of Sciences (PNAS)

• Forecasting Open-Weight AI Model Growth on HuggingFace (arXiv, 2025)

PI: Dr. Pin-Yu Chen, IBM; Dr. Jianxi Gao, RPI

Under Review @ Nature Machine Intelligence

• 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

Under Review @ Nature Communications

Published 

• CRAFT: Training-Free Cascaded Retrieval for Tabular QA (ACL, 2026)

Co-Author: Adarsh Singh | PI: Dr. Jianxi Gao, RPI; Dr. Soham Dan, Microsoft; Dr. Vivek Gupta, Arizona State University
Association for Computational Linguistics (ACL)

• Exploring the Robustness of Language Models for Tabular QA via Attention Analysis (TMLR, 2025)

Co-Author: Sixue Xing | PI: Dr. Soham Dan, IBM; Dr. Jianxi Gao, RPI

Transactions on Machine Learning Research 2025

• Convolutional Neural Network-based EEG Emotion Classification with the Forward Selection Wrapper technique for Channel Selection (JSTOR, 2020)

Advisor: Dr. Jon G. Sigurjonsson, University of Iceland

JSTOR 2020