Analyzing LLM Instruction Optimization for Tabular Fact Verification

Instruction optimization provides a lightweight, model-agnostic approach to enhancing the reasoning performance of large language models (LLMs). This paper presents the first systematic comparison of instruction optimization for tabular fact verification, a task that requires numerical, compositional, and schema-grounded reasoning over structured data. We evaluate four prompting paradigms, direct prediction, Chain-of-Thought (CoT), ReAct with SQL tools, and CodeAct with Python execution, across three benchmarks (TabFact, PubHealthTab, SciTab) and two model families. Using the DSPy framework, we study three optimizers: COPRO, MiPROv2, and SIMBA. Instruction optimization consistently improves verification accuracy, with MiPROv2 yielding the most stable gains for CoT, and SIMBA providing the largest benefits for tool-augmented agents, particularly at larger model scales. Behavioral analyzes reveal that SIMBA encourages more direct reasoning paths by applying heuristics, which enhances numerical comparison in CoT reasoning and helps avoid unnecessary tool calls in ReAct agents. Comparing different pipelines, CoT remains effective for tabular fact checking, especially with smaller models. ReAct agent built with larger models can achieve competitive performance but requires careful instruction optimization.