When Hiring Becomes Agent-Mediated: Evaluating Access and Recurrence in Two-Agent R\'esum\'e Screening
cs.AI · cs.CL
Quick verdict
Claims two-agent résumé screening reallocates, not just expands, access vs one-call; confounded procedure contrast and small cohorts cap confidence.
Summary
The paper asks whether replacing one-call LLM résumé screening with a two-agent exchange—an employer-side agent probing requirements and a candidate-side agent answering only from the résumé—changes who advances to human review. On 600 constructed résumé–job pairs (200 LinkedIn postings × strong/borderline/weak strata) evaluated with GPT-5.5 and Claude Opus 4.7, two-agent screening advances more applications (33.3%→39.3% GPT-5.5; 34.0%→35.5% Opus 4.7), but changes decisions in both directions, ruling out uniform relaxation. Across three runs on a 191-pair borderline pool, pass-instance rates rise from 4.5% to 26.2% and from 6.5% to 16.1%. Even with one-call cutoffs re-read to match two-agent pass volumes, persistent pass sets differ (Jaccard .436 and .333), and no cutoff sweep reconstructs the two-agent set. In fresh reruns of discovery-selected cases, two-agent-only selections recur less often than shared ones (11/20 vs 34/36), significant for GPT-5.5 (p=.0021) but not Opus 4.7 (p=.161), while one-call reruns show no comparable decline. Limitations, mostly acknowledged: the procedures differ simultaneously in compute, decision rules, and a two-agent-only skill gate, so no component is isolated; GPT-5.5 generated the résumés it also evaluates; cohorts are small and persistence-selected; no code or data release is stated.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Two-agent screening changes screening decisions in both directions rather than uniformly relaxing one-call decisions (Table 1: GPT-5.5 removes 6 strong and adds 42 borderline; Opus 4.7 moves both strata both ways).
- At matched pass volumes the two procedures advance substantially different applications, and no threshold on the stored one-call mean reconstructs the two-agent persistent set (best Jaccard .436 GPT-5.5 / .361 Opus over a full cutoff sweep).
- In fresh reruns, two-agent-only persistent selections recur less often than shared selections (11/20 vs 34/36), significant for GPT-5.5 (p=.0021) but not Opus 4.7 (p=.161), while one-call reruns show no comparable decline (31/35 vs 29/36).
- Aggregate pass rates are insufficient evidence of procedural equivalence; evaluating agent-mediated screening requires application-level membership comparison plus direct rerun validation.
Strengths
Re-execution treated as a first-class measurement
Nine runs (A–I) with frozen manifests and endpoints; every Run-A strong-stratum disagreement rerun under both procedures (Appendix B); D–F produced 336 procedure-level records with no selective reruns (Appendix D). This is far beyond the single-shot evaluation typical of the area.
Rate matching plus a full cutoff sweep preempts the threshold objection
Section 4.2 re-reads stored one-call scores at .590/.575 to match two-agent volumes, and Appendix C sweeps every distinct stored cutoff, showing max persistent-set Jaccard of .436 (GPT-5.5) and .361 (Opus 4.7) — the non-recoverability claim does not hinge on a cherry-picked threshold.
Deterministic, outcome-independent testbed validation
Unusually disciplined inference language
Reusable evaluation framework
Weaknesses
Fully confounded end-to-end contrast with no ablation
Missing non-compensatory one-call readout
Headline recurrence contrast rests on small, asymmetrically selected cohorts
Generator–evaluator dependence on a synthetic testbed
Reproducibility artifacts absent
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks a procedure-level question: when résumé screening moves from a single LLM call over a résumé–job pair to a two-agent exchange (an employer-side agent probing requirements; a candidate-side agent answering only from the résumé), does who reaches human review change, and do those changes recur? The authors explicitly do not claim either procedure makes better hiring decisions. The operative claims are: (1) two-agent screening changes decisions in both directions (Table 1: GPT-5.5 removes 6 strong and adds 42 borderline applications; Opus 4.7 moves both strata in both directions); (2) at matched pass volumes the procedures advance different applications, and no one-call cutoff recovers the two-agent persistent set (best Jaccard .436 GPT-5.5 / .361 Opus 4.7 across a full cutoff sweep, Appendix C); (3) procedure-unique selections recur less in fresh executions — clearly for GPT-5.5, less certainly for Opus 4.7. Claims are scoped to "the evaluated procedures," and the paper consistently disclaims attribution to dialogue, role separation, or compute.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Does a two-agent résumé screening system pass more candidates than a single-prompt LLM check?
In this study, yes: the two-agent exchange advanced more applications (from 33.3% to 39.3% with GPT-5.5 and from 34.0% to 35.5% with Claude Opus 4.7, on 600 constructed résumé–job pairs). But it also rejected some applications the single call passed, so the change is a reallocation of access, not just a looser threshold.
Is a two-agent hiring pipeline more repeatable than a single LLM call?
The paper's evidence is mixed. Overall fresh rerun persistence was similar for both procedures (80.4% vs 84.5%), but applications selected only by the two-agent procedure recurred less often than applications both procedures selected (11/20 vs 34/36) — clearly for GPT-5.5 (p=.0021) and not significantly for Opus 4.7 (p=.161). The study does not establish which procedure makes better hiring decisions.
What data was used, and were real applicants involved?
No real applicants were involved. The testbed contains 600 constructed résumé–job pairs built from 200 LinkedIn job postings in three fit strata (strong/borderline/weak), checked by a deterministic non-AI scorer. Because there are no downstream hiring outcomes, the study cannot say whether either procedure is more accurate or fairer.
Is the code or dataset from this paper available?
The paper text contains no statement releasing code, prompts, run logs, or the generated testbed, and early runs lacked logged model identifiers, so exact replication currently depends on the authors' artifacts.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 1 is available only as an inline text rendering (19/24/12 and 16/12/8 memberships); visual layout cannot be independently verified, though the stated Jaccards are arithmetically consistent.
- Prompts, decision-layer code, and run logs are not part of the extracted text, so prompt design and rule implementation could only be judged from their prose description (Appendix A).
- The text appears complete (Sections 1–6 plus Appendices A–D), but the actual PDF layout and any supplementary material beyond it could not be checked.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.