Does a model's stated reason for rejecting a candidate do any work?
cs.CL · cs.AI · cs.LG
Quick verdict
Tests whether an LLM's stated rejection reason is causal by inserting the named fact; results mixed and fragile, but unusually honest with strong validation.
Summary
The paper asks whether a language model's stated reason for rejecting a candidate — usually a fact its profile lacks — actually influences its choice. Under greedy decoding, the experimenter inserts a real corpus sentence stating the named fact into the rival's profile and re-asks, against a length-matched irrelevant control at the same profile and both sentences at an unmentioned third option: a 2×2 design of content against location (conditions R0–R4) with eight deterministic integrity gates. Three runs on 2WikiMultihopQA cover six open models: Part A (n=387), Part B (n=314, fresh checkpoints), Part C (n=563, six options). Results are mixed: the content effect at the named rival clears correction only in Part A (OR 3.57, Holm p=0.0210, robust to single-model exclusion), failing Part B's replication and Part C. The designed contrast — content at the unmentioned option — does not clear correction in Part A (Holm p=0.2428). The strongest result is a pure placement effect with no content claim (Holm p=0.0008). Free-text choices and single-token probabilities disagree in direction on location contrasts. Validating the string-rule pipeline caught eight defects, including a parser that read the rejected option as chosen in 17.1% of responses. Fluency and related confounds mean the content effects are bounded, not established; the six models reduce to about three lineages.
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
- Supplying the model-named missing fact at the named rival's profile moves the discrete choice more than a length-matched irrelevant control in Part A (OR 3.57, Holm p=0.0210), surviving every single-model exclusion.
- The designed contrast — the same content effect at an option no model mentioned — does not clear correction in Part A (Holm p=0.2428) and replicates on only one model in Part C.
- Placement alone moves the choice: the identical irrelevant sentence is more effective at the named rival than at a third option (Part C Holm p=0.0008), so every content result must be checked against its location's disruptiveness.
- The free-text choice read and the forced single-token probability read agree in sign on all six content cells but disagree in direction on four of six location cells, and three candidate explanations for the disagreement find no support.
- Validating the deterministic string-rule pipeline caught eight defects; the uncorrected choice parser would have reported six surviving contrasts instead of four.
Strengths
Judge-free falsifiable design
A rejection names a checkable text claim; the test inserts the named fact and re-reads the choice under greedy decoding, with an R0 determinism check (0 flips across 1,264 rows, §3.6). Figure 1 walks a real item through all five conditions.
Multiplicity and robustness discipline
All twelve discrete contrasts live in one Holm family (Table 3); every key contrast reports exact p under each single-model exclusion (A R1−R2: 0.0118/0.0072/0.0347; C R2−R4 likewise), plus Cochran's Q with an explicit low-power caveat (§4.1).
Instrument validation as a first-class contribution
Negative-result transparency
Full artifact release
Weaknesses
The headline content effect fails replication
Confounds matched only post hoc
Probability measure under non-random missingness
Mechanism ambiguity is unresolvable in this design
Thin effective model diversity and selection skew
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This workshop paper (LLM4XAI @ CIKM 2026) tests a minimal causal question about LLM self-explanations: when a model rejects a candidate while naming a fact the candidate's profile lacks, does that stated fact *do any work*? The falsification condition is checkable without a judge — insert a real corpus sentence stating the named fact (§3.3), re-ask under greedy decoding, read whether the choice flips. Claims are scoped with unusual care in Table 1: the content effect at the named rival is "supported" only in Part A (R1−R2, OR 3.57 [1.54, 8.26], Holm p=0.0210); the designed content contrast at the unmentioned option is explicitly *not* supported in Part A (p=0.2428); the strongest surviving result is a content-free placement effect (C R2−R4, p=0.0008). The author's own framing — content contrasts "bound an effect rather than establish one" — is the accurate summary of the evidence.
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
What does this paper test about LLM explanations?
It tests whether a language model's stated reason for rejecting a candidate — typically naming a fact the candidate's profile lacks — actually influences its subsequent choice. The experimenter inserts a real corpus sentence stating that named fact into the profile and re-asks the same question under deterministic (greedy) decoding, comparing against matched irrelevant-sentence and wrong-location controls.
Did the models' stated rejection reasons turn out to be causal?
Partially and inconsistently. Supplying the named fact at the profile the model itself named moved the choice more than an irrelevant control in the largest run (odds ratio 3.57, Holm-corrected p=0.0210), but the effect failed to replicate on a second model roster, and it weakened below nominal significance when repair and control sentences were closest in fluency. The author concludes the contrasts bound an effect rather than establish one.
Is code or data available for this paper?
Yes. The paper releases MIT-licensed code, per-item records containing every raw model response, and seeded analysis scripts on GitHub (ArchitRastogi20/contrastive-rejection-test), with a README mapping every reported number to the script that regenerates it.
What was the biggest measurement problem the author found?
During validation, the string rule that parsed the model's chosen option returned the option the model had just rejected in 17.1% of adjudicable responses. Uncorrected, this parser error would have changed the paper's conclusions from four surviving contrasts to six; all reported numbers use the corrected parser, verified against a ground-truth rule on 10,038 rows.
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 2 is assessed via its fully tabulated numbers in Table 3; the figure itself was not visually inspected.
- The abstract is truncated mid-sentence in the extraction, but the full body text covers the same content.
- Review is based on the extracted text only; the released repository's contents were not independently inspected.
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.