Hướng dẫnAI evaluationmodel driftrepeated measurementresearch reproducibilitystatistical uncertainty

Did an AI Model Upgrade Change Your Research Results? Audit Drift with Frozen Tasks and Paired Retests

A result change after a model upgrade may come from a real capability change, random generation, a different prompt, a changed evaluator, or a changed service configuration. This guide provides a reproducible audit using frozen tasks, repeated runs, task-level paired comparisons, and separate thresholds for average quality and critical failures.

Author: WeJot

Cập nhật gần nhất 29 phút đọc

When AI is used to draft survey questions, code open-ended answers, probe in interviews, or summarize reports, it becomes part of the measurement process rather than a neutral accessory. If the model, system prompt, retrieval corpus, safety policy, or service implementation changes, the same input can produce a different result. A single comparison before and after an upgrade cannot establish model-version drift, and a public leaderboard cannot substitute for a test on the tasks that matter to the study.

The practical answer is to treat the AI call as an experiment. Freeze the tasks and operating conditions, repeat the calls, compare the same tasks across versions, and assess average quality separately from critical failures. The U.S. National Institute of Standards and Technology identifies changes over time, fine-tuning, drift, and decay as risks that require continuing monitoring. Its generative-AI profile also asks organizations to evaluate performance under conditions similar to deployment and to set explicit acceptance thresholds. The version number is therefore only a trigger for review; the real question is whether the distribution of outcomes on our tasks has changed, where it changed, and whether the change crosses a defensible use threshold.

A score moving from 80% to 75% is not yet evidence

Generative evaluation is not a single exam with one fixed answer sheet. Biderman et al., drawing on years of maintaining a language-model evaluation harness, document how sensitive results are to evaluation setup. The exact prompt format, even down to whitespace, can change performance, and the best prompt style varies by model. Their review also warns that single-run point estimates hide uncertainty and recommends reporting variance across runs or other statistical evidence. Two percentages from two isolated runs therefore mix version effects with sampling noise.

The prompt is not the only variable. A system role, output schema, temperature, token limit, tool configuration, retrieval material, content filter, batch size, or provider-side model alias can change without the product name changing. NIST similarly treats third-party system changes, fine-tuning, drift, and decay as matters that require disclosure and monitoring. An audit record must capture enough of the running environment to reconstruct what was actually tested.

Public leaderboards are a discovery signal, not a project result. Singh et al. found that private model variants, selective score disclosure, and unequal sampling can distort a prominent preference leaderboard. Standardized holistic evaluation programs address the same problem by testing models on shared scenarios and metrics and by publishing raw prompts and completions. A leaderboard shift may justify retesting your workflow, but it does not prove that your coding, interviewing, or questionnaire task changed in the same direction or by the same amount.

Freeze a task set that belongs to the workflow

The central asset in a drift audit is a task set that remains unchanged before and after the upgrade. Tasks should come from real work and include both common and costly failure cases. A useful starting structure is roughly 60% high-frequency scenarios, 25% ambiguous or boundary cases, and 15% mandatory correctness or safety gates. Those proportions are a design heuristic, not a universal law. For open-ended coding, add sarcasm, multiple themes, mixed languages, and empty answers. For interview probing, add refusals, emotional escalation, missing facts, and leading premises. Freeze this task set together with data version and lineage records instead of selecting new examples at every retest.

Each task also needs a fixed scoring standard. Structured extraction can be assessed on field completeness, correctness, and traceability. Open-ended coding should use a frozen codebook, rules for multiple codes, and an explicit “insufficient evidence” category. An interview guide should be judged on whether it asks for concrete experience, repeats a claim rather than an experience, or turns a guess into a fact. If another model serves as judge, freeze its model, prompt, version, and scoring examples, then manually review a sample. Otherwise an apparent regression may simply be a different judge.

Snapshot fieldWhat to retainWhy it matters
TaskInput, task ID, scenario, and human reference standardKeeps before and after on the same questions
ModelModel identifier, call date, published version or endpoint, and change notesSeparates a planned upgrade from an alias change
ParametersSystem prompt, temperature, token limit, stop rules, tools, and retrievalRules out a changed operating condition
OutputRaw answer, safety block, latency, and error codeSupports human review and failure attribution
JudgeAutomated grader version, prompt, thresholds, and human reviewPrevents the standard from drifting with the model

Repeat runs, but do not turn them into extra respondents

Running the same task three, five, or more times helps estimate generation variance, but those outputs are not independent study participants. The task remains the analytical unit. First summarize each task across repetitions, for example as a pass rate or a majority result with an odd number of runs, and then compare the same task before and after the upgrade. Treating five generations of 120 tasks as 600 independent observations shrinks the standard error artificially and can turn unstable variation into a false claim of significance.

Miller frames an evaluation as sampling from an unseen super-population of questions. The recommendations include resampling answers to reduce conditional variance, performing inference on paired task-level differences when comparing models, and planning the number of tasks around power and the smallest meaningful difference. Sample size therefore begins with representative tasks, not merely repeated calls on one easy prompt. More repetitions on a narrow task cannot repair a narrow task set.

A worked version comparison

Suppose an open-ended coding assistant is retested quarterly. The team freezes 120 real tasks. Each task is run three times under the old and new versions, and a majority of the three runs determines whether that task passes. The following counts are entirely illustrative: 78 tasks pass under both versions; 24 pass only under the old version; 10 pass only under the new version; and 8 fail under both.

Paired resultTasksInterpretation
Pass under both78Stable success
Old only24Candidate evidence of regression
New only10Candidate evidence of improvement
Fail under both8Persistent failure requiring separate diagnosis

The old version passes 102 of 120 tasks, or 85.0%; the new version passes 88, or 73.3%. The paired difference is an 11.7 percentage-point decline. An exact McNemar test on these illustrative counts gives a two-sided p of about 0.024, driven mainly by the 24 tasks that changed from pass to fail versus the 10 that changed in the opposite direction. That does not establish that the new version is worse on every Chinese open-coding task. The claim is limited to the frozen tasks and scenario coverage sampled here.

The next step is subgroup diagnosis. If the decline is concentrated in negation, multiple intents, or quotation-based evidence, while simple single-theme coding remains stable, the actionable finding is a loss of stability on complex semantic coding, not a generic statement that the model became worse. If the output scores remain stable while service errors rise on one day, the immediate issue is availability, timeout, or retry behavior, not model-quality drift.

Decide with average quality and critical failures separately

ObservationBest interpretationAction
Overall difference is within repeat-run noise and critical failures do not riseNo stable drift evidenceRetain the version and continue scheduled monitoring
Overall decline is statistically clear but remains inside the pre-set operating thresholdChange exists; impact is currently boundedStart with a small rollout, extra review, and a rollback switch
Average is stable, but failures rise in health, legal, financial, or employment tasksCritical-failure structure changedStop automated use in those contexts and return to human handling
Correctness is stable, but repeated answers vary much moreStability declinedReduce automatic publication and add retries or human adjudication
Automated judging and human review point in opposite directionsThe rubric or judge driftedPause the version decision and recalibrate the standard

This triage converts “is the new model better?” into auditable questions: Did average quality cross a threshold? Which task types changed? Are the worst failures still acceptable? Are repeated runs stable? Do automated and human judgments agree? A one- or two-point change may be worth recording for low-risk writing support. The same average change can still require suspension when it appears in consent, sensitive-data redaction, or medical information routing. The higher the consequence of failure, the less acceptable it is to let a pooled average hide a local breakdown.

A compact drift-audit checklist

  • Write the use contract first: state what the AI may do, what must be human-confirmed, and what may enter a downstream decision.
  • Freeze tasks and standards: retain inputs, codebooks, scoring rules, and a human-reviewed sample before the upgrade.
  • Save the run snapshot: record model identifier, date, parameters, system prompt, tools, retrieval material, and error codes.
  • Repeat within each task: use repetition to estimate generation variance, while keeping the task as the analytical unit.
  • Start with the paired table: report tasks that improved, regressed, passed in both, and failed in both before presenting the overall difference.
  • Stratify by scenario: report ordinary, boundary, and critical tasks separately instead of relying on one average.
  • Pre-set release rules: define the smallest meaningful difference, maximum critical failures, and rollback trigger before seeing the new result.
  • Keep a human exit: high-risk contexts, low-confidence outputs, judge conflicts, and abnormal service errors need a stop rule.

Where this method stops

A drift audit compares AI task performance across changes in the operating environment. It does not show that a product redesign improved user experience and cannot replace real respondents’ reactions to a survey or interview. Population shares, treatment effects, and satisfaction still require a real sample, an appropriate research design, and outcome measures. The audit also cannot prove that one internal model parameter caused a change, because providers may alter post-training, safety policy, infrastructure, and routing at the same time. The method identifies outcome-level drift and its limits, not the vendor’s internal cause.

If the task standard itself is unstable, invest in human agreement, a codebook, or cognitive pretesting before running the model many times. Repeated evaluation of an ill-defined rubric merely disguises a measurement-design problem as model instability. Likewise, a public leaderboard change can justify a retest, but it cannot support the claim that your open-ended coding improved by the same amount.

In a research workflow such as WeJot, AI can reduce drafting and first-pass screening effort, but version audit remains tied to the actual task. Keep a frozen set of coding examples, compare repeated runs, review the dimensions that carry decision risk, and give the research owner the evidence needed to decide whether to expand use. Recording the model version alongside the questionnaire version, cleaning rules, weights, and analysis scripts is part of making a study reviewable, reproducible, and interpretable. Final disclosure should follow the same traceable approach described in the research methodology report guide.

Method and source basis

The method and risk requirements above draw on NIST AI 600-1, the Generative AI Profile (2024); Biderman et al., Lessons from the Trenches on Reproducible Evaluation of Language Models (arXiv:2405.14782); Miller, Adding Error Bars to Evals (arXiv:2411.00640); Singh et al., The Leaderboard Illusion (arXiv:2504.20879); and Bommasani et al., Holistic Evaluation of Language Models (DOI 10.1111/nyas.15007). The 120 tasks and output counts are entirely illustrative rather than empirical benchmarks. The statistical test demonstrates paired reasoning; a real project must plan its task mix, repetition structure, and smallest meaningful difference for its own workflow.

References

  1. https://doi.org/10.6028/NIST.AI.600-1
  2. https://arxiv.org/abs/2405.14782
  3. https://arxiv.org/abs/2411.00640
  4. https://arxiv.org/abs/2504.20879
  5. https://doi.org/10.1111/nyas.15007

Tìm hiểu thêm

Những nội dung này có thể hữu ích cho bạn.

Quay lại Knowledge Base

Vẫn chưa giải quyết được vấn đề của bạn?

Hãy cho chuyên gia tư vấn biết mục tiêu nghiên cứu của bạn, chúng tôi sẽ cùng bạn sắp xếp các bước tiếp theo.