A Comparative Review of LLM Architectures in AI Resume Generators: LLaMA vs. Proprietary Models

In the modern job market, the resume has become a critical document, a bridge between a person's history and a future opportunity. Yet, for many, especially those whose careers have taken non-traditional paths or who have performed essential but uncredited work, translating their experience into the standard language of industry is a difficult task. To solve this, developers have begun building artificial intelligence tools designed to write these resumes automatically. These tools rely on large language models, which are computer systems trained on vast amounts of text to understand and generate human language. Think of these models as highly sophisticated engines that can read a messy, personal story and rewrite it into a polished, professional format. However, there is a fundamental split in how these engines are built and operated. Some are open-source, meaning their inner workings are public and can be run on a computer owned by the user, while others are proprietary, meaning they are closed systems owned by a company and accessed only through the internet. This distinction matters deeply because it determines who controls the data, how much the tool costs to run, and how well it can be tailored to specific needs.

A researcher at the University of Wuppertal, Schmidt Albert, set out to understand the real-world implications of this split by examining the architectural choices behind these AI resume generators. Rather than building a new tool or testing one directly, the researcher conducted a structured review of existing documentation. The study focused on two main examples: the RealCV system, which uses the open-source LLaMA model to convert academic curricula vitae into industry resumes, and the various proprietary systems that rely on commercial models like GPT. The goal was to map out the trade-offs between these two approaches across five specific areas: where the data lives, how the tool can be customized, the cost of running it, how transparent the system is, and its ability to handle unusual or "invisible" work experience.

The review found that the choice between open-source and proprietary models is not a simple matter of one being better than the other, but rather a choice between different kinds of control and convenience. When it comes to data control, the open-source approach offers a significant advantage for privacy. Because systems like RealCV can be run on a local server, the sensitive personal information in a resume never leaves the organization's own computers. In contrast, proprietary models require sending data to an external company's servers to be processed. While these companies have privacy policies, the data still travels outside the user's direct control, which can be a barrier for organizations with strict security rules.

Customization presents another clear divide. Open-source models can be fine-tuned, a process where the model is given additional training on specific data to learn the exact language and needs of a particular industry. This allows for deep, specialized adjustments. Proprietary models, on the other hand, are generally customized through prompt engineering, where the user gives detailed instructions to the model without changing its underlying code. While this is easier to do and requires less technical skill, the study suggests it may not achieve the same level of deep specialization as fine-tuning.

Cost and scalability also pull in opposite directions. Running an open-source model requires an upfront investment in computer hardware and the expertise to maintain it, but once that is in place, the cost to process thousands of resumes can be very low. Proprietary models have no upfront hardware cost; users simply pay a fee for every document processed. This makes them attractive for small-scale use, but the study notes that the costs can add up quickly for large organizations processing high volumes of applications.

Transparency is perhaps the most stark difference. With open-source models, the researchers can see exactly how the system works, including the data it was trained on and the specific settings used. This allows for independent verification and builds trust. Proprietary models are often described as "black boxes," where the internal details are kept secret by the company. While these companies claim to have strong safety measures, the lack of public documentation makes it impossible for outsiders to fully audit how the system makes its decisions.

Finally, the study looked at how well these systems handle "invisible experience"—the kind of work that doesn't fit a standard job title, such as mentoring, community organizing, or caregiving. The RealCV system, built on the open-source LLaMA model, was explicitly designed to recognize and translate these non-traditional experiences into valuable industry skills. The documentation for this system is clear about this capability. In contrast, the documentation for proprietary models does not specifically address this task, leaving it unclear whether they can perform the same translation without further testing.

The paper concludes that neither approach is universally superior. The best choice depends entirely on the specific needs of the user. If an organization prioritizes data privacy, deep customization, and the ability to audit the system, the open-source path offers the necessary tools. If the priority is ease of use, low upfront cost, and immediate access to powerful language capabilities without technical maintenance, the proprietary route is more suitable. The study highlights that the field is still maturing, with significant gaps in how these systems are documented, particularly regarding the costs of running them at scale and the specific methods used to customize them. By laying out these trade-offs clearly, the research provides a framework for developers and organizations to make informed decisions about which architectural path to take as they build the next generation of career tools.