Authors: Swara Sandip Lende, Professor Bangar Abhishek Popat, Professor Bhosale S.B.
Abstract: Artificial intelligence (AI) is increasingly used to accelerate scientific discovery, from predicting protein structures and molecular properties to extracting knowledge from the scientific literature. A large share of experimental evidence in papers is reported in numerical tables, yet automated analysis of such tables remains difficult. This paper presents a focused narrative review of verified sources on AI for scientific discovery, concentrating on large language model (LLM) based understanding of scientific text, scientific tables and numerical reasoning. It examines three connected themes: knowledge extraction and question answering over scientific text, table question answering benchmarks, and reasoning methods (direct prompting, chain-of-thought, program-of-thought, self-consistency and graph-based reasoning). The literature shows that scientific table benchmarks have moved from mostly lookup tasks towards a 1,607-question benchmark that requires multi-step numerical reasoning; on it, GPT-4.1 mini reached only 27.0 exact match under direct prompting on the computation-and-language subset, chain-of-thought prompting raised this to 71.5, and accuracy still fell from 74.6% on small tables to 59.2% on large ones. However, the evidence is concentrated in computer science tables, depends on LLM-generated and LLM-filtered questions, and agreement among human verifiers was only moderate (Fleiss' kappa 0.55). The review contributes a comparative analysis of benchmarks and methods, a proposed conceptual architecture linking extraction, cell selection, stepwise reasoning, verification and expert review, and a prioritised list of research gaps.
