acceptodds
Under review as a conference paper at ICLR 2027

Shared Autoregressive Context Can Distort Relationships in Synthetic Data

Abstract

Large language models can generate several records within one autoregressive completion, making earlier answers available as context for later records. This paper shows that such shared-completion batching can distort relationships among variables in the resulting synthetic data, using controlled tests on synthetic survey respondents. In a matched experiment on European Social Survey profiles, generating ten rather than one respondent per request increases mean absolute error in within-country correlations by 48-58% for Qwen3.8-27B and 114-127% for Llama-3.3-70B-Instruct across three seeds, holding profiles, examples, questions and decoding parameters fixed. The distortion primarily reflects exaggerated relationship strength, while retaining substantial agreement with the human ordering of correlations. Controlled interventions establish answer history as a causal channel: re-pairing the same preceding values, with profiles and marginal distributions fixed, changes correlations among subsequently generated responses. Hiding preceding answers reduces correlation error in the tested settings but worsens marginal accuracy. Exploratory corrections across social-attitude, health and economic data likewise show that lower correlation error can coexist with worse marginal distributions and regression estimates. Request construction is therefore part of the data-generating process, and synthetic-data validity must be evaluated against the analyses the generated data are intended to support.

Primary area:
generative models
Keywords:
Large language models, Synthetic data, Survey simulation, Generative-model evaluation, Silicon sampling, Request batching

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_HzZXOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how generating several survey respondents in one completion affects synthetic data. Generating ten respondents together produces larger correlation errors than generating them separately, mostly by making the relationships stronger than those in human data. The paper also tests how earlier answers affect later responses. Hiding earlier answers or correcting the data can reduce correlation error, but can make marginal distributions or regression estimates worse.

Critical strengths
  • I like the clear experimental setting: a comparison of the same profiles under one-person and ten-person requests while keeping the examples, questions, and decoding settings fixed.

  • The experiment that changes the pairings of earlier answers is useful. The values stay the same, but the relationships among later generated responses change, showing that answer history matters.

  • I also like that the paper checks correlations, marginal distributions, and regression estimates together. Improving one metric does not necessarily make the synthetic data better for other analyses.

  • The authors run additional checks with different temperatures, prompt wording, and output constraints, and report where the results are less consistent.

Critical weaknesses
  • The writing is difficult to follow. Terms such as “relationship strength,” “pattern agreement,” and “dependence fidelity” sound similar, but their differences are not clearly explained. Sentences and captions also contain too many experimental details and qualifications. It is often hard to tell what each experiment is trying to show.

  • It is still unclear why earlier answers make the relationships stronger, rather than simply changing them. More explanation of what the model might be doing would help readers understand the finding.

  • The paper shows that one-person generation and corrections can improve correlation error while making other measures worse. But it is unclear what readers should do with this result. A concrete example of how batching changes an analysis would make its importance clearer.

  • The main comparison uses one or ten respondents per request. It does not show how the effect changes between these settings, or how many preceding respondents are needed for the distortion to appear.

Questions for authors
  • Does the order of respondents matter? What happens if similar demographic profiles are grouped together rather than mixed?

  • Why does batching mainly exaggerate the strength of relationships? Do you think the model follows patterns in earlier answers or generates less individual variation?

  • When hiding history lowers correlation error, how much comes from recovering the human pattern and how much comes from simply making correlations weaker?

  • Can you show an example where batching changes the conclusion of an analysis, in addition to its average error?

  • Given that one-person generation improves correlations but can worsen marginal distributions, when would you recommend generating respondents separately?

1 comment
AuthorsOpenReview ↗

Reply to Reviewer HzZX

Thank you for your careful review and for questions that shaped the new experiments.

Terminology. Section 2 now defines relationship strength, pattern agreement and correlation error with equations, and the paper uses only these terms. Table 1 and Figure 1 show what each experiment changes and what it is meant to show. We have also cut extraneous detail and qualifications from sentences and captions.

Q2. Why relationships become stronger. Each new respondent moves towards the respondents before it, and in a batch this builds up, so respondents spread along one axis. When people differ mainly along one shared dimension, every pair of variables tied to it correlates more strongly. The model therefore follows the pattern in earlier answers, rather than producing less individual variation. Batching multiplies the variance along that axis by 1.8–4.0, while variance in other directions changes much less (0.96–1.22; Section 3.3, Appendix C).

How many preceding respondents are needed. One is enough. Two-person requests are already worse than one-person requests in every seed. Error rises by 0.023 for Qwen and 0.085 for Llama with two-person requests, and by 0.094 and 0.223 with ten-person requests. Within a request, error also rises with position (Section 3.1, Figure 3A).

Q1. Order and grouping. Reversing the order within requests changes error little. Grouping similar profiles lowers error for Qwen and raises it for Llama, in every seed (Section 3.3).

Q3. Recovery or weakening when hiding history. The gain comes mostly from weakening. For Qwen, masking brings correlation error back to about its one-person level (Table 8). There is some recovery: the error left after rescaling the correlations falls from 0.081 to 0.065, and the Pearson agreement rises from 0.848 to 0.894. The rank ordering of pairs, however, barely changes (ρ\rho from 0.770 to 0.782). For Llama, the ordering gets worse (ρ\rho from 0.798 to 0.616). These results are from NHANES (Section 3.3, Table 7).

Q4. An example. In an exploratory regression of trust in police on income, happiness, satisfaction with democracy and age, the coefficient on satisfaction with democracy is 0.36 in the human data. Batching inflates it in all three models, for example from 0.34–0.37 to 0.49–0.61 for Llama. Signs and the ordering of predictors mostly survive, so batching changes how strong a relationship looks rather than whether it exists (Section 3.5).

Q5. When to generate separately. We recommend generating respondents separately whenever an analysis uses relationships, such as correlations or regressions, because batching inflates them and an instruction does not prevent this. Batching gives better marginals in our data, but through the same process, so it is not better data in general (Section 5).

We would be glad to answer any further questions before the discussion closes on 18 November.

Reviewer_8c4SOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper shows that simulating multiple survey respondents in a shared completion (also referred to as batching) can distort relations between variables in the generated responses. Specifically, compared with generating one respondent at a time, batching generally exaggerates variables' correlations but yields more aligned variables' marginal distributions. It also investigates the effects of re-pairing and hiding preceding answers in the same batch, and explores correction approaches.

Critical strengths

Respondent batching is commonly used in LLM-based simulations, and understanding its effects on the simulation fidelity is a practically important and interesting problem. The paper investigates this problem on three survey datasets, and conducts controlled experiments to explore the potential causes of misalignment, which provides useful quantitative evidence to deepen our understanding.

Critical weaknesses
  1. The writing of the paper is difficult to follow, which strongly hinders understanding of the experimental setup and interpretation of the results. In section 3, the experimental setups are not clearly described, including the motivations, procedures, controls, and the datasets used. In the controlled experiments of re-pairing and hiding previous answers, the changes in the setup and the evaluation metrics are not explained clearly. The accuracy metrics in section 2 are also described only in words, which can be difficult to understand without their mathematical definitions (which are currently deferred to the appendices). The exploratory correction procedure in section 4 is also not clearly explained, with a lot of important details deferred to the appendices.

  2. The paper gives limited additional insights beyond the direct experimental findings. While the experiments in section 3.2 explore the effects of changing pairings and hiding histories as two potential causes, these do not yet reveal deeper or more systematic insights into why they cause misalignment this way, or whether other factors also contribute to the misalignment. Thus, while the experimental findings are useful, they still give limited deeper understanding of when batching causes misalignment, and how to alleviate such misalignment.

Questions for authors

None

4 comments
AuthorsOpenReview ↗

Reply to Reviewer 8c4S

Thank you for your review. We agree that the writing hid the experimental logic, and so we have substantially rewritten the paper.

Setups and metrics. Section 2 now defines every measure with an equation and gives the error of a real sample (0.019) as a benchmark. Table 1 lists each experiment's question, what it changes, what it measures, and its data and models. Section 3.3 explains the re-pairing and masking experiments, and Appendix B describes them step by step. The correction procedure was exploratory and is no longer part of the main argument. We summarise this in Section 4 and defer a longer discussion to Appendix F.

Why batching misaligns. This question led to our main new results (Section 3.3), and we are grateful for it.

  • New respondents move towards the ones before them. When we shift the earlier answers, the next respondent follows 0.31–0.57 of the shift with one predecessor and 0.76–0.90 with nine.
  • The effect builds up within a request. The first respondent is undistorted, and error rises with position.
  • This spreads respondents along one axis, which makes relationships stronger and marginals wider at the same time. Our mechanism explains part of the effect, but not all of it (Appendix C).

Section 3.3 also covers the other factors you mention, such as the order and grouping of profiles, and Section 3.4 covers where assimilation comes from.

How to alleviate it. Telling the model to generate respondents independently does not help. Masking the earlier answers removes much of the penalty but does not restore the human pattern. Writing real respondents' answers in as the earlier records does lower error, to 0.045 for Qwen against 0.170 for one-person requests (Table 2). However, it pulls each generated respondent away from the person's own answers, so it suits population-level estimates only. For analyses of relationships, researchers should generate respondents one at a time (Section 5).

We would be glad to answer any further questions before the discussion closes on 18 November.

Reviewer_8c4SOpenReview ↗

Reply to Author Rebuttal

I thank the authors for the responses and revision. The newly added metric definitions in section 2 and the experiment overview in table 1 of section 3 improved the clarity of presentation.

However, I find the newly added subsections 3.3 and 3.4 quite dense and hard to follow. It seems that almost every paragraph in the two subsections has its own experiment (or experiments), and the specific setup and purpose of each experiment, as well as how the different experiments relate to each other, are barely explained.

While the authors offer assimilation as an explanation of misalignment of batching, I am not fully convinced by the empirical evidence. For example, the two experiments in the first two paragraphs of subsection 3.3 vary preceding responses along principal components (why principal components, instead of other directions?) and examine how changes in later responses relate to changes in preceding responses. These do demonstrate dependence between earlier and later responses, but are not directly connected to the discrepancy between real and synthetic correlations.

I will maintain my score of 2.

Thank you for reading the revision so carefully, and for saying that Section 2 and Table 1 helped. In the next version we will set out each experiment in the same order (question, manipulation, what is held fixed, result), and add a short map at the start of Section 3.3 showing how they connect.

On why we shift along the leading principal component: this is the direction in which batching changes the data. Ten-person requests multiply the variance along this axis by 1.8–4.0, while variance in other directions changes much less (0.96–1.22; Section 3.3, Appendix C). Variables that load on a shared axis become more correlated when respondents spread out along it. Shifting earlier answers along this axis therefore tests the direction most relevant to the inflated correlations, rather than an arbitrary one.

We agree that the shift experiments show dependence between earlier and later respondents, and do not by themselves explain the gap between real and synthetic correlations. We do show two results that make the link, but you are right that both are partial (see Section 3.2, Table 2 and Appendix C, Table 9). This is why we describe assimilation as a mechanism that explains part of the effect, not all of it. Your comment shows that the paper should state this chain more directly, and we will do so.

Thank you again for engaging with the revision.

I thank the authors for the further clarification. I will maintain my current score.

Reviewer_TMWEOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper asks whether generating several synthetic survey respondents in one LLM completion changes the relationships among variables in the output. In matched ESS experiments, ten-person requests raise within-country correlation error by roughly 50 to 130 percent compared with one-person requests. The authors also show that changing or hiding earlier answers shifts later correlations, and that post-hoc corrections can lower correlation error while making marginal distributions and regression estimates worse.

Critical strengths

The question matters for applied work, since many studies already generate respondents in batches.

The matched design is careful, and the authors report validity failures openly rather than dropping them.

The finding that correlation accuracy and marginal accuracy can pull in opposite directions is useful for anyone building synthetic data.

Critical weaknesses
  • One- and ten-person requests differ in output length, format, and context, so the effect cannot be attributed to sharing alone. The title and abstract nevertheless present it as a property of shared completion.
  • Whole-request rejection removes all ten respondents when one answer is invalid, and invalid rates differ across arms. The paper never bounds the resulting selection bias.
  • Correlations estimated from 100 respondents per country carry substantial sampling error, but the paper does not show what error a resampled human sample would produce, so the reported magnitudes have no reference point.
  • The history interventions change both the pairing of values and the token positions. The language of a “causal channel” is stronger than this evidence supports.
  • ACS results are largely invalid or unstable across seeds, yet the generalization claims are broad.
  • Qwen and Llama use different precision and serving setups, so direct cross-model comparisons are unreliable.
  • Several correction analyses were designed after the original results were known and reuse previously evaluated references. The headline “better correlations, worse marginals” finding is therefore not confirmatory.
  • Key results are spread across many appendices and reported on different scales, which makes the main contributions hard to follow.
Questions for authors

Can you separate the effect of sharing from changes in format and context, for example with a condition that keeps the output structure identical?

What correlation error does a resampled human reference produce at 100 respondents per country?

How much do the main contrasts change when rejected ten-person requests are bounded under best-case and worst-case assumptions?

3 comments
AuthorsOpenReview ↗

Reply to Reviewer TMWE

Thank you for your detailed and specific review. We agree with several of your points.

Q1. Sharing versus format. We now separate these two (Section 3.2, Table 2). We keep the ten-person prompt and format and regenerate one respondent at a time. Only the source of the earlier answers changes: the same completion, independent one-person answers, or real answers. With answers from the same completion, the model reproduces ordinary ten-person error (largest gap 0.014). Sharing accounts for 69–75% of the penalty for Qwen and 82–84% for Llama. For phi-4 it accounts for 41–44%, so format and context matter more for this family. The abstract and Section 3.2 now say this. We kept the title because it refers to the shared context of a request, including its format, and says only that this context can distort relationships. If you prefer, we would be glad to retitle the paper "Generating Several Records in One Completion Can Exaggerate Relationships in Synthetic Data".

Q2. A resampled human reference. Drawing 100 real respondents per country 1,000 times gives an error of 0.019, with 95% of draws between 0.016 and 0.023. The batching increases are five to twelve times larger. The value is low because each correlation pools all 2,000 respondents, and Section 2 now explains this.

Q3. Rejected requests. Filling rejected respondents with their own one-person answers, or with other accepted answers, changes the contrast by at most 0.014 (0.017 in the replication runs). Under an extreme fill that puts every answer at a random end of its scale, eleven of twelve contrasts stay positive. The exception is the Qwen run with the most rejections (97.5% valid). Section 3.1 reports these bounds.

Pairing and token positions. Across the eighteen orders, each value appears equally often in each position, although a single re-pairing also changes the order of values (Appendix B.1). An exploratory check within orders leaves most of the contrast. We agree that "causal channel" overstated this experiment. The revision drops the phrase, and the new history-source and shift experiments (Sections 3.2–3.3) carry the causal argument.

ACS. We agree. ACS is now a stress test, because fewer of its outputs are valid and Llama's effect changes sign across seeds. Our claims about correlation error rest on ESS and NHANES.

Cross-model comparisons. We now compare conditions only within a runtime, and we do not attribute differences in size to the models. The main result replicates when both models are rerun at FP8 on different hardware.

Post hoc corrections. We agree. They are now in Appendix F and labelled exploratory. The marginal half of the trade-off is, however, now confirmatory: batching improves marginal error in every seed of all three models, for example from 0.142 to 0.110 for Llama (Section 3.5).

Presentation. Correlation error is now the single main measure, always compared with the 0.019 of a real sample, and Table 1 lists every experiment. The new experiments also show one way sharing produces the penalty (Section 3.3).

We would be glad to answer any further questions before the discussion closes on 18 November.

I would like to thank the authors for their response. I acknowledge that I have read the response and will maintain my initial assessment.

Thank you for reading our response and, again, for the very helpful review. Could you tell us which concerns remain open after the revision, in particular whether Table 2 (Q1), the 0.019 sampling reference (Q2) and the rejection fills (Q3) address your questions? If the remaining concern is presentation, we would be glad to know which parts are still hard to follow.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of revisions

We thank the reviewers for careful and constructive reviews. All three valued the matched design but found the writing hard to follow. The paper has therefore undergone a substantial rewrite, both to present our existing claims and argument more clearly, and to address the substantive issues raised by each reviewer.

The two most important changes are Table 1 (p. 4), which lists every experiment and the question it answers, and Table 2 (p. 6), which separates sharing a completion from the request format.

These revisions should make the paper's three claims much clearer:

  1. Batching exaggerates relationships. Ten-person requests raise correlation error by 48–58% for Qwen3.8-27B and 114–127% for Llama-3.3-70B. These increases are five to twelve times the error of a real sample of the same size (0.019). The result replicates in new runtimes and in phi-4, and it grows with request size.
  2. Much of this comes from sharing a completion. With the prompt, position and format held fixed, answers from the same completion account for 69–75% of the penalty for Qwen and 82–84% for Llama. For phi-4 they account for 41–44%, so format matters more there (Table 2).
  3. A mechanism: assimilation that compounds. Each new respondent moves towards the respondents before it. With nine predecessors, it follows 0.76–0.90 of a shift we apply to them (Figure 3). Because the earlier respondents were pulled in the same way, the effect builds up within a request and spreads respondents along one axis. This inflates correlations and widens marginal distributions, and an instruction to generate respondents independently does not stop it. This mechanism explains part of the effect, not all of it (Appendix C).

What has not changed. The question, design, data and headline estimates are unchanged. The new experiments added in Sections 3.1–3.3 answer specific requests in the reviews. Section 3.4, on where assimilation comes from, extends our answer to the question of why. The main new material is Table 1, Figure 1, Sections 3.2–3.4, and the replication, request-size and rejection results in Section 3.1. Sections 2, 3.5, 4 and 5 are revised.

Other changes. We have defined each measure with an equation in Section 2, collected the caveats in the limitations section, added a recommendation on when to batch (Section 5), moved the exploratory corrections to an appendix, narrowed the ACS and "causal channel" claims, and shortened the appendix from 43 pages to 12.

How we ran the new experiments. Each followed a dated protocol written before its data were generated, on open-weight models run locally. The supplement lists every prediction and decision rule, and the dated protocol files are available to the AC on request. Appendix C lists every prediction and its outcome, including those that failed, and labels the exploratory and post hoc analyses. A second comment below summarises further results.

AuthorsOpenReview ↗

Further results: where assimilation comes from, and a larger model

Reviewers 8c4S and HzZX asked why models behave the way we document, and Section 3.4 reports what we found:

  • It is present before post-training. In OLMo-3-7B, whose training checkpoints are public, assimilation is already strong at the end of pretraining (0.70) and falls with later training (0.39 after preference tuning). Qwen3.5-9B also assimilates slightly less after post-training.
  • It depends on who the earlier respondents are, not on how their answers are written. Rewriting the earlier answers as prose, or shuffling their fields, leaves most of the effect. Relabelling the earlier respondents as coming from a distant country reduces it by more than a third.
  • Clustered training data are enough to produce it. Small transformers trained from scratch assimilate only when the records in their training documents share a group, and then with the strength that is optimal for their training data.
  • A larger model shows the same penalty. In Qwen3.8-Flash-Next (180B parameters), ten-person requests raise correlation error by 58–71% (two seeds), and answers from the same completion account for 79% of the penalty (one seed). Fewer of its outputs are valid (91.5–94.0% of ten-person requests), but filling the rejected respondents with plausible answers still leaves an increase of 0.050–0.078 (Appendix C).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.