Overview

This chapter isolates whether non-qualifying self-promotion can move an LLM’s résumé ranking. In a controlled task, DeepSeek-V3.2 and GPT-4o-mini rank ten synthetic candidates for an IT Support Specialist role, comparing baseline and injected versions across homogeneous five-year pools and heterogeneous pools with five- and ten-year experience. Descriptive and instructive insertions are placed near the top or bottom, with randomized order and repeated rounds. Rank gain and success rate measure movement, not shortlist or hiring probability. The results vary by model and wording: DeepSeek-V3.2 is highly responsive in homogeneous pools, while GPT-4o-mini is more resistant to descriptive language but more responsive to instructive text. Stronger quality differences reduce average effects without eliminating occasional boundary-relevant inversions. The study demonstrates a conditional ranking mechanism, not a general production success rate.

Learning outcomes

Rank gain measures movement against a baseline; it does not equal shortlist, interview, or hiring probability.
Inserted language was most influential among similarly qualified candidates, with model and wording producing different responses.
Quality differences reduced average manipulation effects but did not remove occasional rank inversions near a decision boundary.

Keywords

résumérankinginjectionDeepSeek-V3.2GPT-4o-minithreshold

When Self-Promotion Moves the Ranking

To see whether machine-only qualifications are merely an integrity problem or can alter model behavior, the controlled study asks a narrower question: if an inserted statement adds no substantive qualification, does an LLM still move that résumé upward? The answer depends on how similar the candidates are, how the insertion is phrased, and which model is doing the ranking. The experiment is useful because it isolates ranking influence from the real-world prevalence question, allowing readers to interpret a rank change without mistaking it for a guaranteed shortlist or hire.

A controlled contest between baseline and injected rankings

The researchers construct a fixed résumé-ranking task around an IT Support Specialist job description. In each round, a model receives ten candidate résumés and must return a strict ordering from strongest to weakest fit. Every candidate receives a different rank, and the order in which the résumés are presented is randomized. This matters because a model might otherwise favor the first or last résumé it reads, making a position effect look like an injection effect.

The résumés are synthetic and share a common template. Candidate quality is deliberately simplified to make the comparison interpretable. The main homogeneous condition gives every candidate five years of relevant experience. The heterogeneous condition gives five candidates ten years and five candidates five years. All candidates meet the job description’s five-year minimum, so the comparison is between qualified applicants with a controlled difference above the threshold, not between a qualified and an obviously ineligible résumé. An appendix repeats the homogeneous analysis with ten years for every candidate and reports the same qualitative pattern.

The study tests two forms of inserted self-promotion. A descriptive insertion makes an evaluative statement about the candidate. An instructive insertion goes further by directing the model to classify the candidate as superior regardless of the résumé’s content. The researchers vary whether the text is placed near the top or near the bottom, randomizing that position once per round and applying it consistently to the injected résumés in that round. This setup does not reproduce the full range of real document concealment; it isolates how a ranker responds to additional language.

Two models are evaluated under the same ranking task: DeepSeek-V3.2 and GPT-4o-mini. Single-injection conditions use 100 independent rounds, while multi-injection conditions use 30 rounds because they require substantially more computation. The analysis in this chapter begins with one injected résumé at a time. That comparison is the cleanest test of whether a rare, strategically inserted signal can differentiate an otherwise similar candidate; the interaction among several injectors changes the question and will be considered separately.

The study uses two outcome measures. Rank gain compares the candidate’s position in a baseline ranking with its position when the résumé contains the insertion. Formally, the baseline rank minus the injected rank is positive when the candidate moves upward, negative when the candidate moves downward, and zero when the position is unchanged. Success rate is the fraction of rounds in which the injected résumé achieves a strictly positive rank gain.

Published by DeepMox

FAQ

How were ranking changes measured in the controlled résumé study?

The study compared each candidate’s baseline rank with the rank after inserted text; baseline rank minus injected rank is rank gain, and the share of rounds with positive gain is success rate. These measures describe model ordering, not a hiring outcome.

Why can a lower-quality résumé gain more positions but succeed less often?

A lower-quality résumé has more room to move when an insertion works, so its average gain can be larger even if it moves upward in fewer rounds. A stronger candidate may start higher and therefore have a higher success rate with smaller gains.

What limits applying this ranking experiment to real hiring?

The results come from synthetic ten-candidate pools, one IT Support Specialist role, two models, controlled decoding, and two insertion styles; they do not measure larger workflows, live systems, defenses, or downstream decisions.

Previous11 minAutomated Résumé Screening: Hidden Qualifications

Next10 minAutomated Résumé Screening: Rare Signals and Saturation

Automated Résumé Screening: When Self-Promotion Moves Rankings

When Self-Promotion Moves the Ranking

A controlled contest between baseline and injected rankings

FAQ