How Reproducible Are Evaluation Conclusions? A Self-Audit of LLM-Inferred Prompt Structure

Hugging Face Daily Papers Papers
llm-evaluation reproducibility prompt-structure self-audit evaluation-audit

Summary

This paper audits the reproducibility of LLM evaluation conclusions, finding that small-sample evaluations are unstable and can mislead, and it recommends reporting additional metrics for reliability.

Evaluations of LLM systems routinely average over small prompt sets and report models as a ranked table. We ask how much confidence such a table deserves, using LLM-based prompt-structure inference as the case study: eight open model variants across five families and 8B to 675B parameters, caching disabled, 293 raw intermediate representations persisted. The measured phenomenon is unstable to begin with. Identical calls do not reliably recover identical structure, with mean node-set Jaccard from 0.39 to 0.96 and 72% of prompt-model cells never node-set-perfect. Auditing the evaluation weakens its conclusions further, and this is our main contribution. Under a joint cluster bootstrap over prompts, only the bottom of the ranking is firm: the two least reproducible models hold rank in 99% and 86% of replicates, the middle four in 27% to 48%, and the top two in 68% each, so the table identifies the worst model reliably but does not reliably identify the best. Two equally defensible rules for merging repeated campaigns change four of eight rows and move the study-wide headline by 7 percentage points. Checking the inferred structure against ground-truth annotations shows reproducibility cannot be read as accuracy. And four of the eight endpoints were withdrawn within ten weeks of measurement, so the study as specified can no longer be run. Small-sample LLM evaluations can therefore look far more definitive than their evidence supports. We recommend reporting rank stability, per-cell provenance, executed sensitivity comparisons, raw per-run outputs, and a measurement date alongside any ranking.
Original Article
View Cached Full Text

Cached at: 09/29/26, 08:12 AM

Paper page - How Reproducible Are Evaluation Conclusions? A Self-Audit of LLM-Inferred Prompt Structure

Source: https://huggingface.co/papers/2609.30074

Abstract

EvaluationsofLLMsystemsroutinelyaverageoversmallpromptsetsandreportmodelsasarankedtable.Weaskhowmuchconfidencesuchatabledeserves,usingLLM-basedprompt-structureinferenceasthecasestudy:eightopenmodelvariantsacrossfivefamiliesand8Bto675Bparameters,cachingdisabled,293rawintermediaterepresentationspersisted.Themeasuredphenomenonisunstabletobeginwith.Identicalcallsdonotreliablyrecoveridenticalstructure,withmeannode-setJaccardfrom0.39to0.96and72%ofprompt-modelcellsnevernode-set-perfect.Auditingtheevaluationweakensitsconclusionsfurther,andthisisourmaincontribution.Underajointclusterbootstrapoverprompts,onlythebottomoftherankingisfirm:thetwoleastreproduciblemodelsholdrankin99%and86%ofreplicates,themiddlefourin27%to48%,andthetoptwoin68%each,sothetableidentifiestheworstmodelreliablybutdoesnotreliablyidentifythebest.Twoequallydefensiblerulesformergingrepeatedcampaignschangefourofeightrowsandmovethestudy-wideheadlineby7percentagepoints.Checkingtheinferredstructureagainstground-truthannotationsshowsreproducibilitycannotbereadasaccuracy.Andfouroftheeightendpointswerewithdrawnwithintenweeksofmeasurement,sothestudyasspecifiedcannolongerberun.Small-sampleLLMevaluationscanthereforelookfarmoredefinitivethantheirevidencesupports.Werecommendreportingrankstability,per-cellprovenance,executedsensitivitycomparisons,rawper-runoutputs,andameasurementdatealongsideanyranking.

View arXiv pageView PDFGitHub0Add to collection

Get this paper in your agent:

hf papers read 2609\.30074

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2609.30074 in a model README.md to link it from this page.

Datasets citing this paper1

#### dipankarsarkar/llm-evaluation-self-audit Viewer• Updatedabout 19 hours ago • 1.69k

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.30074 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Can LLMs Write Reliable Rubrics? A Meta-Evaluation for Experiment Reproduction

arXiv cs.CL

This paper presents the first systematic meta-evaluation of LLM-generated rubrics for reproducing experiments from research papers. It reformulates rubrics into a checklist format and evaluates generation settings both intrinsically (semantic similarity) and extrinsically (score alignment), finding that augmented settings improve downstream evaluation alignment but generated rubrics are often overly fine-grained and biased toward high scores.

When the Judge Changes, So Does the Measurement: Auditing LLM-as-Judge Reliability

arXiv cs.CL

This paper audits the reliability of LLM-as-judge evaluation by showing that changing the evaluator model can shift scores even when candidate responses are fixed, and it examines scaling and upgrade paths for Qwen3 and MiniMax models, concluding that judge upgrades are not interchangeable and proposing best practices for reporting.