Reasoning over complex scientific tables is an essential yet largely underexplored topic. Existing benchmarks often fail to capture the structural and content challenges of complex scientific tables, such as irregular layouts, multi-row headers, and symbolic notations. To address these challenges, we introduce SciTableQA, a benchmark comprising 320 scientific tables in 5 scientific domains, and 8,730 human-verified QA pairs categorized into two types of reasoning tasks. Our tables cover 8 structural features and 4 content features. The questions include 4 interrogative and 3 imperative question styles. To answer the questions, the model needs to reason across up to 14 rows and 13 columns of a table. We evaluate the performance of three widely used Large Language Models (LLMs)—GPT-3.5-turbo, Llama 3-8B, and Mistral-7B against the benchmark dataset. Our evaluation focuses on reasoning accuracy, cross-LLM generalization, and explanation validity. Our findings reveal that arithmetic tasks are more challenging than cell selection tasks. LLMs’ performances on both types of tasks vary across domains. Compared with existing TableQA benchmarks, we believe SciTableQA provides a more challenging dataset for evaluating machine learning models’ reasoning capability on complex scientific tables. Our dataset is available at https://huggingface.co/datasets/Kehindeajayi01/SciTableQA