Reasoning Models Ace the CFA Exams

TL;DR

The paper provides a comprehensive, reproducible evaluation of state-of-the-art reasoning models on a 980-question mock CFA dataset spanning Levels I–III. It compares baseline LLMs to leading reasoning models (e.g., GPT-5, Gemini 3.0 Pro) under zero-shot and chain-of-thought prompting, reporting near-perfect Level I performance and strong Level II/III results, including a CRQ score of 92.0% for Gemini 3.0 Pro. It discusses curriculum updates, data validity, prompt design, and automated scoring limitations, concluding that modern reasoning models have surpassed prior CFA benchmarks and offer unified baselines for future research. The findings highlight the models’ capacity for high-level synthesis and numeracy in finance, while noting persistent challenges in Ethical Standards and the importance of robust evaluation protocols. Overall, the work signals advancing capabilities in automated financial reasoning and provides a rigorous framework for benchmarking future CFA-level reasoning systems.

Abstract

Previous research has reported that large language models (LLMs) demonstrate poor performance on the Chartered Financial Analyst (CFA) exams. However, recent reasoning models have achieved strong results on graduate-level academic and professional examinations across various disciplines. In this paper, we evaluate state-of-the-art reasoning models on a set of mock CFA exams consisting of 980 questions across three Level I exams, two Level II exams, and three Level III exams. Using the same pass/fail criteria from prior studies, we find that most models clear all three levels. The models that pass, ordered by overall performance, are Gemini 3.0 Pro, Gemini 2.5 Pro, GPT-5, Grok 4, Claude Opus 4.1, and DeepSeek-V3.1. Specifically, Gemini 3.0 Pro achieves a record score of 97.6% on Level I. Performance is also strong on Level II, led by GPT-5 at 94.3%. On Level III, Gemini 2.5 Pro attains the highest score with 86.4% on multiple-choice questions while Gemini 3.0 Pro achieves 92.0% on constructed-response questions.

Figures (6)

• Figure 1: Sample mock CFA exam questions by level. Cases are shown in blue, questions in red, and answer choices in green. Examples are illustrative and not actual exam content.
• Figure 2: Example of a Concept Misapplication error, where the model incorrectly selects between two related propositions.
• Figure 3: Example of a Rule Application error, where the model misapplies ethical standards to a specific case vignette.
• Figure 4: Example of a Misinterpretation of Evidence error, where the model incorrectly flags a normal portfolio activity as a sign of a poor benchmark.
• Figure 5: Example of a Concept Oversimplification error, where the model provides a common but incorrect generalization instead of the nuanced answer.