Overview

This chapter tests how competitive intensity changes ranking effects when several résumés carry the same non-qualifying signal. In ten-candidate synthetic pools, increasing the number of injected résumés reduced average rank gain and success rate; under the tested homogeneous DeepSeek-V3.2 conditions, success approached zero once roughly 80% or more of résumés were injected. That is saturation, not immunity: a rare signal can still separate one candidate from similar peers, and a lower-quality résumé can occasionally outrank a stronger peer near a consequential boundary even when average robustness looks strong. The analysis compares no-injection, rare-injection, and high-intensity conditions, while noting limits from two models, one IT-support role, narrow experience-based quality, and no downstream shortlist outcomes. Robustness therefore must be tested across competition, quality, model, prompt, and operational cutoff.

Learning outcomes

Saturation reduces the relative advantage of an inserted signal as more candidates carry it; it does not prove immunity.
Rare manipulation remains the more exposed tested regime because an unusual signal can separate otherwise similar candidates.
Mean rank gain can hide cutoff risk, so robustness tests should inspect boundary crossings across competition and quality conditions.

Keywords

résuméscreeninginjectionsaturationcompetitionthreshold

The Advantage of Being Rare—and the Cost of Saturation

The controlled ranking results change meaning when more than one candidate manipulates the same screening task. This chapter focuses on competitive intensity: why an inserted signal can help when it is rare, lose its differentiating value when many résumés carry it, and still create fairness-relevant distortions near a decision boundary. The practical judgment is conditional rather than categorical: widespread manipulation may reduce the average advantage of any one insertion, but that does not make a ranking system safe when a small number of candidates can still move ahead of stronger peers.

When a differentiating signal becomes ordinary

In a homogeneous pool, every résumé begins with the same controlled level of experience. With only one injected résumé, the inserted language can act as a distinguishing cue. With several injected résumés, that cue is no longer unique. The model must rank multiple candidates who all contain the same kind of extra signal, so the advantage that once separated one résumé from its peers is competed away.

The multi-injection experiments make that intuition measurable. As the number of injected résumés increases in the ten-candidate pool, both average rank gain and success rate decline. Under DeepSeek-V3.2, the decline is monotonic in the tested homogeneous conditions: the early advantage shrinks toward zero, and success approaches zero once roughly 80% or more of the résumés are injected. GPT-4o-mini is more robust under descriptive injection and shows smaller gains, but the same diminishing-return pattern appears under the instructive variant.

This is saturation, not immunity. A candidate can benefit from an unusual signal precisely because other candidates do not have it. Once the signal becomes common, it provides less information for separating candidates. The model may still read or respond to the inserted language, but its relative ranking value falls because the same cue appears across the pool. A low average gain at high injection intensity therefore says that manipulation has stopped differentiating candidates in that condition; it does not say that the model has learned to recognize the content as illegitimate or that all downstream risk has disappeared.

The distinction between absolute and relative effects is important. A model could assign favorable language some influence for many candidates while still giving none of them a systematic advantage over the others. In a ranking task, what matters is not only whether a score changes but whether one candidate’s change is larger, more reliable, or better aligned with the cutoff than a competitor’s. Saturation reduces that relative separation. It does not necessarily mean that the extra text has vanished from the model’s reasoning or that the model has correctly identified the text as outside the intended evidence.

The same logic appears in heterogeneous pools, where high- and low-quality candidates differ in experience. The experiment confines multiple injections to one quality group so that it can observe competition within that tier. Under descriptive injection, the benefits are limited and short-lived for both groups. Under instructive injection, low-quality candidates can show larger initial gains when only a few of them inject, but those gains decay as more candidates in the same tier do so.

Published by DeepMox

FAQ

What is saturation in multi-injection résumé screening?

Saturation occurs when the same inserted signal becomes common enough that it no longer differentiates one candidate from the others. Rank gain and success rate can then decline, but saturation does not mean the model rejects the content or that downstream risk has disappeared.

Why should ranking tests compare rare and widespread manipulation?

They should compare both because a rare insertion can create a relative advantage, while widespread insertion tests whether competition erodes that advantage. A no-injection baseline is also needed to show how the model ranks substantive qualifications without the extra signal.

Can average ranking robustness hide boundary risk?

Yes. A low average gain can coexist with an occasional large rank inversion near a top-K or review cutoff, where a small movement may change who receives human attention. The experiment does not quantify those downstream outcomes.

Previous10 minAutomated Résumé Screening: When Self-Promotion Moves Rankings

Next12 minAutomated Résumé Screening: Restoring Evidence

Automated Résumé Screening: Rare Signals and Saturation

The Advantage of Being Rare—and the Cost of Saturation

When a differentiating signal becomes ordinary

FAQ