Language models such as GPT and Llama have shown remarkable ability on diverse natural language tasks, yet their performance on complex table tasks (e.g., NL-to-Code, data cleaning, etc.) continues to be suboptimal.To improve their performance, task-specific fine-tuning is often needed, which, however, require expensive human labeling and is prone to over-fitting.In this work, we propose TABLE-SPECIALIST, a self-trained fine-tuning paradigm specifically designed for table tasks.Our insight is that for each table task, there often exist two dual versions of the same task, one generative and one classification in nature.Leveraging their duality, we propose a Generator-Validator paradigm to iteratively generate-then-validate training data from language models, to finetune stronger TABLE-SPECIALIST models that can specialize in a given task, without using manually-labeled data.

Tables and table-based use cases play a crucial role in many important real-world applications, such as spreadsheets, databases, and computational notebooks, which traditionally require expert-level users like data engineers, data analysts, and database administrators to operate. Although LLMs have shown remarkable progress in working with tables (e.g., in spreadsheet and database copilot scenarios), comprehensive benchmarking of such capabilities remains limited. In contrast to an extensive and growing list of NLP benchmarks, evaluations of table-related tasks are scarce, and narrowly focus on tasks like NL-to-SQL and Table-QA, overlooking the broader spectrum of real-world tasks that professional users face. This gap limits our understanding and model progress in this important area.In this work, we introduce MMTU, a large-scale benchmark with over 28K questions across 25 real-world table tasks, designed to comprehensively evaluate models ability to understand, reason, and manipulate real tables at the expert-level. These tasks are drawn from decades’ worth of computer science research on tabular data, with a focus on complex table tasks faced by professional users. We show that MMTU require a combination of skills -- including table understanding, reasoning, and coding -- that remain challenging for today's frontier models, where even frontier reasoning models like OpenAI GPT-5 and DeepSeek R1 score only around 69% and 57% respectively, suggesting significant room for improvement. We highlight key findings in our evaluation using MMTU and hope this benchmark drives further advances in understanding and developing foundation models for structured data processing and analysis.Our code and data are available at https://github.com/MMTU-Benchmark/MMTU and https://huggingface.co/datasets/MMTU-benchmark/MMTU.

Modern data stores increasingly rely on metadata to enable diverse activities such as data cataloging and search. However, metadata curation remains a labor-intensive task, and the broader challenge of metadata maintenance—ensuring its consistency and usefulness—has been largely overlooked. In this work, we tackle the problem of resolving relationships among metadata concepts from disparate sources. Inferring these relationships are critical for creating clean and consistent metadata repositories, and a central challenge for metadata integration. We propose OpenForge, a two-stage prior-posterior framework for metadata integration. In the first stage, OpenForge exploits multiple methods including fine-tuned large language models to obtain prior beliefs about concept relationships. In the second stage, OpenForge refines these predictions using the Markov Random Field, a probabilistic graphical model. We formalize metadata integration as an optimization problem, where the objective is to identify the relationship assignments that maximize the joint probability of assignments. The MRF formulation allows OpenForge to capture prior beliefs while encoding critical relationship properties, such as transitivity, in probabilistic inference. Experiments on four datasets show the effectiveness and efficiency of OpenForge. In a use case of matching two metadata vocabularies, OpenForge outperforms GPT-4, the second-best method, by 25 F1 points.

Existing table search techniques define table relatedness with unionablility and/or joinability. While these are valuable, they do not suffice for most data analysis tasks that involve numerical data, which is often aggregated over geographical, temporal, or other groups. In this demonstration, we showcase ARTS, a novel table search system centered on the unique concept of aggregate relatedness. By leveraging pre-trained language models, ARTS offers a superior column semantics understanding capability, with good labels created for both textual and numerical columns. This demonstration will offer attendees hands-on interaction with our system, revealing its potential in effectively addressing real-world data analysis challenges.

Exploratory data analysis can uncover interesting data insights from data. Current methods utilize "interestingness measures" designed based on system designers' perspectives, thus inherently restricting the insights to their defined scope. These systems, consequently, may not adequately represent a broader range of user interests. Furthermore, most existing approaches that formulate "interestingness measure" are rule-based, which makes them inevitably brittle and often requires holistic re-design when new user needs are discovered. This paper presents a data-driven technique for deriving an "interestingness measure" that learns from annotated data. We further develop an innovative annotation algorithm that significantly reduces the annotation cost, and an insight synthesis algorithm based on the Markov Chain Monte Carlo method for efficient discovery of interesting insights. We consolidate these ideas into a system. Our experimental outcomes and user studies demonstrate that DAISY can effectively discover a broad range of interesting insights, thereby substantially advancing the current state-of-the-art.

Chinese word segmentation (CWS) trained from open source corpus faces dramatic performance drop when dealing with domain text, especially for a domain with lots of special terms and diverse writing styles, such as the biomedical domain. However, building domain-specific CWS requires extremely high annotation cost. In this paper, we propose an approach by exploiting domain-invariant knowledge from high resource to low resource domains. Extensive experiments show that our model achieves consistently higher accuracy than the single-task CWS and other transfer learning baselines, especially when there is a large disparity between source and target domains.