An Empirical Comparison of ReAct, Reflexion, Plan-and-Solve, and Tree-of-Thought Planning Strategies on Financial Question Answering and Numerical Reasoning Tasks

Large language model agents increasingly automate reasoning- and decision-intensive financial workflows, yet the comparative effectiveness of competing planning strategies on finance-specific tasks remains unclear. We conduct a controlled empirical comparison of four widely adopted planning strategies --- ReAct, Reflexion, Plan-and-Solve, and Tree-of-Thought --- on four public financial benchmarks spanning multi-step numerical reasoning (FinQA), multi-turn numerical dialogue (ConvFinQA), hybrid tabular-textual question answering (TAT-QA), and long-document question answering (DocFinQA). Using a shared GPT-4o backbone, a common tool set, and a unified evaluation protocol, we measure execution accuracy, exact-match correctness, per-task-type performance, and per-query token cost across three random seeds. Plan-and-Solve offers the best accuracy-per-dollar on purely numerical tasks, delivering a moderate 2.8-point improvement over ReAct on FinQA at roughly one-seventh the token budget of Tree-of-Thought. ReAct with retrieval dominates on long-document DocFinQA, outperforming Plan-and-Solve by 4.1 points. Tree-of-Thought attains the single highest accuracy on the compound-arithmetic subset of TAT-QA (71.4%) but costs 7.2× more tokens per query than Plan-and-Solve. A manual error typology across 400 failures confirms that each strategy repairs a distinct failure class, and that no single strategy dominates all four financial task types. The findings clarify an existing design-space question rather than propose new methodology.