Why Frontier Models Won’t Solve Resume Matching
By Brian DeAngelis, PhD
Co-Founder & Chief Scientist, AdeptID
WAITING ON THE NEXT MODEL
Every few months, a new frontier model arrives with better reasoning, larger context windows, and stronger benchmarks. It’s tempting to believe that matching resumes to job descriptions is simply waiting for the next generation of models. To wait on the next generation of models is to miss the fundamental problem: the information needed to make a good match often isn’t in the resume or job description.
The limitation isn’t the model. It’s the task we’re asking it to perform.
Modern foundation models are already remarkably good at understanding resumes and job descriptions. They can identify skills, summarize experience, compare qualifications, and reason about fit. If the task were simply “understand these two documents,” we’d already be close to solving it.
Hiring isn’t a language understanding problem. It’s a preference learning problem.
WHAT JOB DESCRIPTIONS DON’T TELL YOU
The fundamental issue is that job descriptions rarely specify what a hiring manager actually wants from a candidate. Job descriptions are often written by committee, recycled from previous openings, padded with “nice-to-have” qualifications, or missing the characteristics that ultimately drive hiring decisions. Recruiters must dissect from the vague to wishful job descriptions and their understanding of the hiring manager’s desires, what the role actually requires and which candidates are most likely to succeed. This leads to why two recruiters working the same role may prioritize completely different attributes. A single recruiter alone may have multiple acceptable candidate archetypes in mind.
A hiring manager looking for a software engineer, for example, may happily hire either a senior distributed systems expert or a fast-growing full-stack generalist. Those candidates may have almost nothing in common on paper, yet both are excellent fits for the role. No model, regardless of how capable, can infer which archetype is preferred if that information is never provided.
That’s why simply replacing GPT-5.6 with GPT-6 doesn’t solve the problem. A more capable model will produce more reasonable guesses, but it’s still guessing. You can’t recover information that was never specified.
THE SOLUTION IS PERSONALIZATION
Rather than asking a single model to make increasingly difficult inferences, recruiting systems should gather more information about what the recruiter, and by proxy the company and hiring manager, actually values. That can come from extracting structured hiring rubrics, asking targeted follow-up questions, learning from historical recruiter and company hiring behavior, or retrieving candidates similar to those previously advanced. Each of these reduces ambiguity in ways that a larger language model alone cannot.
THE RIGHT MODEL FOR EACH STEP
This is also why the best production architectures won’t rely exclusively on frontier models. While a foundation model can perform every step in the pipeline, it’s rarely the right engineering tradeoff. Rubric extraction can be handled by a fine-tuned small language model while candidate evaluation is fundamentally a classification problem. Similar to candidate evaluation, search is best served by embedding models and vector retrieval. Ranking, in turn, improves most when a dedicated model learns from user feedback. Frontier models are most valuable where reasoning and interaction are required, not as a replacement for every component in the system.
Frontier models will continue to improve, and they’ll continue to become an increasingly important part of recruiting software. Nonetheless, the companies that build the best matching systems won’t be the ones using the largest models. Rather, they’ll be the ones that have distinguished themselves as the best at learning exactly what recruiters want before asking a model to make a recommendation.