Bafna Jainit Sushil supervised by Dr. Manish Shrivastava received his Master of Science – Dual Degree in Computational Linguistics (CLD). Here’s a summary of his research work on Advancing Structured Table Understanding: Rubric-Based Evaluationand Multimodal Reasoning

Tables are a fundamental medium for representing structured information across domains such as finance, healthcare, scientific reporting, and enterprise analytics. With the increasing use of large language models and neural systems for table understanding and generation tasks, the reliability of automatic evaluation has become a critical bottleneck. Unlike unstructured text, tables encode meaning through both content and structure, where even minor misalignments in rows, columns, or cell values can significantly alter semantics. However, existing evaluation metrics, largely adapted from natural language generation, fail to adequately capture these structural dependencies and often show poor alignment with human judgment. This thesis addresses the problem of evaluating structured table outputs by proposing TabXEval, a rubric-based and explainable evaluation framework for tables. TabXEval decomposes table evaluation into interpretable criteria covering structural alignment, semantic correctness, and completeness, and employs a two-phase evaluation pipeline consisting of table alignment followed by fine-grained celllevel comparison. By explicitly modeling table structure and content interactions, TabXEval provides diagnostic feedback and achieves strong agreement with human evaluators. To enable systematic assessment of table evaluation metrics, this thesis also introduces TabXBench, a benchmark comprising controlled table perturbations with human-annotated quality judgments. Extensive experiments demonstrate that TabXEval significantly outperforms existing string-based, embedding-based, and LLM-based metrics in terms of sensitivity to structural errors and correlation with human judgment. In addition, this thesis presents MMTABQA, a multimodal table question answering dataset that extends traditional table reasoning tasks by incorporating images within table cells. Evaluations on MMTABQA reveal substantial performance gaps between state-of-the-art models and highlight persistent challenges in multimodal reasoning over structured data. Together, this work advances the evaluation and understanding of structured tables, emphasizing the importance of rubric-based, human-aligned metrics for reliable development and deployment of tablecentric AI systems. 

July 2026