ChatGPT is no longer the only large language model to pass multiple choice sections from CPA Exam practice tests.
After ChatGPT 3.5
These were part of the findings of a
Overall, they found that ChatGPT-4 scored the best, with Claude 3-opus coming in a close second, followed by Google Gemini Advanced, then Mixtral-8x7b-32768. Llama 2B-70b-4096 did the worst.
However, as the results show, not every model did uniformly well on all sections. ChatGPT, while a strong performer overall, was especially good on the BAR section for business analysis and reporting. Meanwhile, although its weakest point is REG, the regulatory area that is mostly devoted to tax regulations, it did better on this section of the exam than any other model. Claude was the best performer in the AUD section on auditing and attestation. While its weakest point was FAR, the section on financial accounting and reporting, even there its performance was second only to ChatGPT. Gemini was the second strongest performer on the BAR section, but did not do so well on REG. Mixtral, overall, had decent enough scores compared to a human but would only pass BAR, making it a mediocre player compared to its peers. Llama was the only one that would not pass any section, and it did especially poorly on REG. It was also the only one that did worse than a human. The average score for human test takers on REG was 59.19%, according to the paper.
"The study revealed that while some LLMs have made significant advances in mimicking the complex decision-making skills required for CPA exams, there remains variability in performance across different sections of the test," said the paper. "This variability underlines the importance of tailored training and specialization in developing LLMs for professional applications such as the CPA exams."
To perform the test, the researchers drew their multiple choice questions from the Becker CPA test preparation suite. Google Gemini, Claude and ChatGPT-4 were accessed via their online platforms. Mixtral and Llama-2b models were accessed through the Groq platform, an advanced computational infrastructure for high-speed AI processing. The questions were directly copied and pasted into the AI platforms from Becker's test preparation material without any additional prompting or modification to ensure each AI model received the questions in their original form as they would appear in a CPA exam context.
Becker's platform randomized the questions in batches of 15 questions, which the research said further mitigated potential selection bias. The tester, responsible for inputting the questions into the AI models, deliberately refrained from reading or evaluating the questions beforehand to prevent any unconscious bias in the prompting process. For each question, the tester selected the AI model's first response marked as "correct," irrespective of any variations in the explanations or outputs provided by different models.
Each AI model was subjected to each multiple choice section of the CPA test three times, allowing for a comprehensive assessment of its performance across multiple attempts. The criterion for determining an AI model's success in this study was achieving a passing score, defined as an average score of 75 or higher, on any given section.
The researchers said the data indicates there is no one universal model for all tasks, so it is important to use the right model for the right applications. For example, the paper concluded that ChatGPT is "the only real option for zero-shot BAR automation," as "no other model came close to its performance, and it had a relatively narrow variance," meaning that ChatGPT-4 could be used to help with automated financial statement preparation or additional forecasting. On the other hand, the researchers said Claude was probably better on auditing-related tasks, which the paper said "is a solid indication that it can be used for fraud detection and internal control validation."
"It is apparent from the results that there is no clear-cut winner," the researchers concluded. "Most companies utilizing AI to perform financial administration functions should use a software infrastructure that allows them to use multiple task-dependent AI models."
However, the researchers did recommend that "model selection for AI in an applied accounting setting should avoid Llama-2B, which performed worse than any other model in every section."
An AICPA spokesperson said that the organization disputes the framing that these LLMs "passed" the exam. It was noted that, in the experiment, retired questions were used; further, results were based on multiple-choice questions only, which represents only 50% of the Exam; and the results were likely scored as a percentage of correct / incorrect, which is not how people score the Exam.
It was also pointed out that the Exam is secure and ChatGPT does not have a role in the test center administration of the CPA Exam for the foreseeable future. Further, live, operational questions are not shared publicly. If they are shared, they are retired. Finally, test question / item statistics, instrumental in scoring the Exam, are not released publicly.
"The AICPA Exams Team, in collaboration with the AICPA BOE (Board of Examiners) are exploring how ChatGPT and AI will impact test preparation, test development, test content, and other areas. The security and validity of the administration of the Exam is of utmost importance to all of us and ChatGPT does not pose a risk to this area for the foreseeable future," said the spokesperson in an email.
