Measuring Validity in LLM-based Resume Screening

1. Problem Statement

The rapid adoption of Large Language Models (LLMs) in resume screening has outpaced rigorous evaluation of their validity. While existing research extensively documents fairness issues (bias against specific demographic groups), it largely overlooks validity: the fundamental ability of a model to correctly identify the most qualified candidate based on job-relevant criteria.

Key challenges in measuring validity include:

• Lack of Ground Truth: Real-world hiring data lacks definitive "correct" answers because human hiring decisions are themselves often biased or inconsistent.
• Train-Test Contamination: Publicly available resume datasets are often used in LLM training, making independent evaluation unreliable due to data leakage.
• Conflation of Metrics: Prior work often conflates fairness (statistical parity) with validity (accuracy), failing to distinguish between a model that is consistently wrong and one that is biased.

2. Methodology: A Ground-Truth-Controlled Framework

The authors propose a novel framework inspired by software engineering testing principles (specifically metamorphic testing and mutation testing) to generate synthetic evaluation scenarios with known ground truth.

A. Conceptual Foundation

The framework defines two types of validity based on psychometrics:

1. Criterion Validity: Relevant qualifications (skills, experience) must positively influence the hiring decision. If Candidate A has strictly more relevant qualifications than Candidate B, the model must prefer A.
2. Discriminant Validity: Irrelevant attributes (demographics, names, unrelated hobbies) must not influence the decision. If two candidates have identical relevant qualifications, the model should treat them as equal (ideally by abstaining from choosing).

B. Synthetic Data Generation Pipeline

To avoid train-test contamination, the authors construct a pipeline that generates resumes from scratch based on live job descriptions:

1. Job Parsing: Extract required ($Q_{req}$) and preferred ($Q_{pref}$) qualifications from job descriptions (sourced from Greenhouse, LinkedIn, Indeed).
2. Base Resume Construction: Generate a "base" resume ($c$) that satisfies exactly the required qualifications.
3. Mutation (Unequal Pairs):
   • Under-qualified ($c^-$): Remove $k$ random required qualifications.
   • Over-qualified ($c^+$): Add $k$ random preferred qualifications.
   • Ground Truth: $c^+ \succ c \succ c^-$.
4. Rewording (Equal Pairs): Generate a reworded version of the base resume ($c_{reword}$) with identical qualifications but different phrasing.
5. Demographic Injection: Append demographic signals to resumes to test discriminant validity. Signals are either:
   • Implicit: Names correlated with race/gender (e.g., "Jamal" vs. "Bradley").
   • Explicit: Awards or organizational affiliations (e.g., "Black Women in Tech Award").
   • Ground Truth: $c_{reword} \sim c$ (Indifference).

C. Evaluation Metrics

The framework evaluates LLMs on pairwise comparisons ($P_j: C \times C \to C \cup \{\perp\}$):

• Criterion Validity Score: The fraction of times the model correctly selects the more qualified candidate.
• Unjustified Abstention: Failing to choose when a clear superior candidate exists.
• Unjustified Selection: Choosing the less qualified candidate.
• Discriminant Validity (Abstention Rate): The fraction of times the model correctly abstains when presented with equally qualified candidates of different demographics.
• Over-Assessment Rate: The rate at which a model selects a candidate from a specific demographic over an equally qualified peer.

3. Key Contributions

1. First Principled Validity Framework: Introduces a systematic method to audit LLM resume screeners for validity (accuracy) and fairness simultaneously, independent of human baselines.
2. Contamination-Free Evaluation: By generating synthetic resumes from live job descriptions, the framework eliminates the risk of models having seen the test data during training.
3. Decoupling Validity and Fairness: Demonstrates that a model can be valid (accurate) but unfair, or consistent but invalid, treating them as distinct, measurable properties.
4. Scalable Audit Tool: Provides a methodology for regulators and companies to perform on-demand, job-specific validity audits (e.g., complying with NYC Local Law 144).

4. Key Results

The authors evaluated 9 frontier models (including GPT-5, Claude Sonnet 4, Gemini 2.5 Pro, Llama 3.3, etc.) across various job roles.

A. Criterion Validity (Accuracy)

• Performance Gap: While larger models generally perform better, validity is far from perfect.
  • Top models (Claude Sonnet 4, GPT-5) achieved ~96-99% validity when differences were large ($k=3$).
  • Critical Failure at $k=1$: When candidates differed by only one qualification (a common real-world scenario), validity dropped significantly (e.g., GPT-4o-mini dropped to 50-59%; Llama models dropped to ~64-73%).
• Error Types: Most errors were Unjustified Abstentions (models refusing to choose even when one candidate was clearly better), rather than selecting the wrong candidate. This suggests models are overly cautious.
• Soft vs. Hard Skills: Models struggled more with "soft skills" (e.g., "passion," "curiosity") and qualifications hidden in resume summaries compared to hard educational credentials.

B. Discriminant Validity (Fairness & Abstention)

• Failure to Abstain: Models frequently failed to abstain when presented with equally qualified candidates.
  • Discriminant validity was generally lower than criterion validity (<0.9 for most models).
  • Models often invented rationales (e.g., "better formatting") to justify choosing one candidate over another despite identical qualifications.
• Demographic Bias:
  • Over-Alignment: Contrary to some prior studies showing bias against minorities, this study found that in forced-choice scenarios, some models (e.g., Claude Sonnet 4) selected Black and female candidates at higher rates than White male candidates when qualifications were equal.
  • Explicit vs. Implicit: Bias was more pronounced when demographic signals were explicit (awards/organizations) compared to implicit (names).
  • Over-Assessment: White men were less likely to be "over-assessed" (selected over an equal peer) compared to other groups.

C. Generalizability

• Results held across different job types (Software Engineering, Nursing, Wind Turbine Technician) and data sources (synthetic vs. real-world resumes from Reddit).
• Performance on real-world resumes was slightly worse than on synthetic ones, suggesting synthetic evaluations are optimistic but directionally accurate.

5. Significance and Implications

• Redefining AI Auditing: The paper argues that "fairness" (statistical parity) is insufficient. A hiring tool must first be valid (able to identify the best candidate). A biased model that is also invalid is doubly dangerous.
• The "Abstention" Trade-off: The study highlights that while abstention is a crucial tool for fairness (preventing arbitrary decisions), current models abstain too often in valid scenarios, potentially rejecting qualified candidates unnecessarily.
• Over-Correction Risk: The finding that some models favor minority candidates in equal scenarios suggests that current alignment techniques (RLHF) designed to mitigate bias may be inducing a new form of invalidity where demographic signals override qualifications.
• Regulatory Impact: The framework provides a concrete, reproducible method for organizations to audit "out-of-the-box" LLMs before deployment, addressing the "black box" nature of commercial AI hiring tools.

In conclusion, the paper establishes that while LLMs show promise in resume screening, they currently lack the robust validity required for high-stakes hiring, particularly in nuanced scenarios with minor qualification differences. It calls for a shift from static benchmarks to dynamic, ground-truth-controlled testing frameworks.