TL;DR
Researchers have developed UCTR-ST, a self-training framework that masters complex reasoning over tables without requiring any human-labeled data. By generating sophisticated synthetic training samples from programs and iteratively refining itself on unlabeled real-world data, the model achieves over 90% of the performance of fully supervised systems.
What Problem Does This Paper Address?
While Large Language Models (LLMs) excel at textual reasoning, structured tabular data remains a hurdle. Tables require understanding spatial relationships, numerical logic, and multi-step operations. Current supervised models are "data-hungry," requiring thousands of manually annotated examples that are expensive and slow to produce.
Furthermore, these models often fail when encountering new topics not seen during training (see Figure 1 below). Existing unsupervised methods usually generate overly simplistic "simple claims" (e.g., just reading a single cell), whereas real-world tasks require "complex claims" involving comparisons, aggregations, and arithmetic across multiple rows and columns.

Methods
The authors propose UCTR-ST (Unsupervised Complex Tabular Reasoning via Self-Training), which consists of three core components:
- Program-Management: Instead of just looking at raw data, the system generates executable programs (SQL queries, logical forms, or arithmetic expressions) and runs them on tables to get ground-truth answers. This ensures the logic is sound and the answers are precise.
- Program-Transformation: To bridge the gap between code and human speech, the framework uses generative models (like GPT-2 and BART) to translate these abstract programs into natural language questions or claims.
- Table-Text Manipulator: Real-world tables often come with surrounding text. The authors designed "Table-To-Text" (splitting a table to create a text context) and "Text-To-Table" (extracting text info into a table format) operators to synthesize hybrid reasoning samples.
The "secret sauce" is the Self-Training loop (illustrated in Figure 3). The model first trains on synthetic data to become a "Teacher." This Teacher then labels unlabeled real-world data. These "pseudo-labeled" samples are merged with the synthetic data to train a "Student" model, which eventually becomes the new Teacher. This iterative process allows the model to adapt to the distribution of real-world language.

Datasets
The framework was evaluated on four diverse benchmarks covering different domains and tasks:
- FEVEROUS: Fact verification using Wikipedia (hybrid table-text).
- TAT-QA: Complex question answering in the finance domain (hybrid table-text).
- WikiSQL: Large-scale question answering on Wikipedia tables.
- SEM-TAB-FACTS: Fact verification for scientific tables.
Key Insights and Ideas
- Programs as Ground Truth: By using programs (SQL/Logic) as the starting point, the model avoids the "hallucination" problem common in pure text generation. If the program executes to "True," the generated claim is guaranteed to be factually grounded in the table.
- Complex vs. Simple Reasoning: The authors highlight that simple row-to-text heuristics aren't enough. As shown in Figure 2, complex claims require the model to understand relationships between multiple cells (e.g., "is 5872 higher than...").
- Bridging the Synthetic-Real Gap: Synthetic data is abundant but often "sounds" different than real human questions. The self-training step is crucial because it allows the model to learn the style and distribution of real-world data without needing the labels for that data.

Key Results
The results demonstrate that UCTR-ST is a powerhouse for low-resource scenarios:
- Unsupervised Performance: UCTR-ST reached 90%+ of the performance of supervised models across various tasks. For instance, on the finance-heavy TAT-QA dataset, it achieved an F1 score of 47.6 without a single human label.
- Few-Shot Superiority: In settings with only 50 labeled samples, UCTR-ST significantly outperformed standard supervised baselines, proving its value as a pre-training or data augmentation tool.
- Ablation Success: The "Table-Text Manipulator" proved vital; models trained with hybrid data (Table ↔ Text) performed significantly better than those trained on tables alone.
Limitations and Future Work
The authors acknowledge that the Program-Transformation module isn't perfect. Occasionally, the generated natural language can lose critical information or contain minor inaccuracies compared to the underlying program.
Future work will focus on broadening the types of reasoning the programs can handle and exploring "auto-program generation" methods that can adapt to new data distributions even more flexibly.
For more details, you can find the code and models at https://github.com/leezythu/UCTR-ST.
Summary updated