Shi-Yu Tian

Shi-Yu Tian

Nanjing University
2 previous affiliations
4Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Shi-Yu Tian focuses on weakly supervised learning, tabular data modeling, and the reasoning capabilities of large language models (LLMs). Their work addresses the challenges of learning from ambiguous or incomplete data, with a particular emphasis on developing robust methods for real-world scenarios where data distributions shift or features are missing. Currently, Tian’s research explores the intersection of mathematical reasoning and tabular data, specifically investigating how LLMs handle complex numerical tasks in the presence of inconsistent information.

Research Trajectory

Tian’s early documented work centered on the problem of label ambiguity within the framework of weakly supervised learning. In 2024, they addressed the challenges of Partial-Label Learning (PLL), a setting where each training instance is associated with a set of candidate labels, only one of which is correct. In the paper CroSel: Cross Selection of Confident Pseudo Labels for Partial-Label Learning, Tian proposed a method to improve the identification of ground-truth labels by leveraging historical model predictions. This work introduced a cross-selection strategy designed to mitigate the noise inherent in pseudo-labeling processes, thereby enhancing the accuracy of identification-based PLL methods.

By 2026, Tian’s research trajectory shifted toward the evaluation and improvement of LLMs, specifically regarding their ability to perform reasoning over structured data. This transition moved from general classification problems toward complex mathematical and logical reasoning. In TabularGSM: Understanding the Limitations of LLMs in Tabular Math Reasoning, Tian investigated the gap between LLM performance on standard math word problems and the requirements of real-world business intelligence applications. This research highlighted the limitations of current models when faced with multi-step numerical reasoning over tables, particularly when those tables contain incomplete or "trapped" information. The work contributed to the field by providing a more scalable and rigorous framework for evaluating tabular reasoning than previous manual collection methods allowed.

During this same period, Tian participated in research regarding the robustness of tabular prediction models in open-environment settings. As a contributor to Handling Tabular Data under Coupled Shifts of Feature Missingness and Distributional Change, Tian assisted in examining the interdependent challenges of column missingness and distributional shifts. This work identified how these two factors can mutually inhibit model performance and proposed strategies for robust prediction when both issues occur simultaneously.

Current Directions

Tian’s current research appears to be concentrated on the intersection of structured data and generative AI. The focus has moved toward identifying and addressing the "traps" that real-world data poses to automated reasoning systems. This includes:

  • Robustness in Tabular Reasoning: Developing benchmarks and methodologies to ensure that LLMs can handle the inconsistencies and missing values typical of industrial and scientific datasets.
  • Weakly Supervised Learning for Structured Data: Applying principles from partial-label learning and pseudo-labeling to improve how models interpret and reason over tables with latent or ambiguous structures.
  • Evaluation of LLM Logic: Moving beyond simple accuracy metrics to understand the specific failure modes of LLMs when performing multi-step calculations and data extraction from tabular formats.

The available evidence suggests a research program that is increasingly concerned with the reliability of machine learning systems in "open-world" environments, where the clean assumptions of laboratory datasets—such as complete features and static distributions—rarely hold.

Representative Work

Profile updated

See something wrong? Report a mistake