The conditional superiority of fast silicon sampling

Nickolas Hock Yuen Lam; Ji Xuan Voo; Xiangyu Ma

Source-linked AI summary

The conditional superiority of fast silicon sampling

Nickolas Hock Yuen Lam, Ji Xuan Voo, Xiangyu Ma

arXiv:2608.14079v1Aug 14, 2026cs.CLcond-mat.mtrl-sci

TL;DR

Silicon sampling may estimate human opinion means well, but evidence remains limited on its fidelity across modes and relational opinion structure. Using Singaporean survey data, this study compares fast and slow sampling and finds fast modes more efficient and at least as faithful, despite broader fidelity limitations.

Problem
Evidence remains limited on silicon sampling’s fidelity beyond American populations and on relational opinion structure beyond item-level and pairwise diagnostics.
Method
The study compares fast and slow silicon sampling against 2,012 Singaporean World Values Survey respondents using geometric and Procrustes analyses.
Results
Fast silicon sampling is more efficient in compute resources and run-time while being monotonically superior to slower modes in algorithmic fidelity.
Takeaways & Limitations
Silicon sampling should be used cautiously as an exploratory instrument rather than a substitute for human respondents in substantive claims about Singaporeans.
Takeaways & Limitations
Silicon samples have limited fidelity to Singaporean respondents, understating opinion variance and distorting covariance and latent contextual structure.

Abstract

from arXiv · show

Silicon sampling can produce surprisingly good population estimates at times. Does doing it fast attenuate such fidelity? In this study, we extend and assess ongoing work in silicon sampling by comparing the algorithmic fidelity of "fast" and "slow" modes of silicon sampling among a nationally representative sample of Singaporean survey respondents. We find that silicon sampling with contemporary frontier models remains a method in early development to be used only with great caution. While silicon samples are able to produce moderately faithful estimates of population means, they continue to understate opinion variance and distort the latent contextual space behind human opinions. Conditional on such limitations, we find "fast" modes of silicon sampling to be relatively superior to traditional "slow" modes of silicon sampling. Fast silicon sampling is significantly more efficient in compute resources and run-time while being monotonically superior to slower modes of sampling in algorithmic fidelity.

1 Introduction

The introduction frames silicon sampling as a promising but still early-stage method and evaluates whether fast sampling offers greater algorithmic fidelity than slow sampling. It reports moderately faithful population means alongside underestimated opinion variance and distorted latent contextual structure, with fast modes relatively superior under these limitations.

Silicon sampling can sometimes recover human subpopulation opinion distributions with surprising accuracy, motivating efforts to create synthetic survey and behavioral panels.
The study evaluates the algorithmic fidelity of fast and slow silicon-sampling modes created by contemporary frontier models.
Silicon samples moderately estimate population means but understate opinion variance and distort the latent contextual space behind human opinions.
Conditional on these limitations, fast silicon sampling is relatively superior to traditional slow sampling.
The paper advances methodological evaluation through geometric and Procrustes analysis of relational fidelity.
It deepens fidelity assessment for an understudied national population, addressing the field’s overwhelmingly American empirical base.

2 Literature review

The literature frames silicon sampling as using LLMs as synthetic surrogates for human survey respondents, while distinguishing slower repeated prompting from faster, more efficient procedures. Existing work favors slow sampling, but whether fast sampling preserves algorithmic fidelity remains unresolved.

Conceptual foundations: Silicon sampling uses large language models as synthetic surrogates for human respondents in public opinion surveys.The approach rests on LLMs’ ability to converse and reason in ways that can resemble human responses.
Sampling modes: Slow silicon sampling repeatedly uses zero- or one-shot prompts to elicit small sets of synthetic responses, whereas fast sampling is defined as a distinct mode.
Prior literature: 66 of 75 preprints use slow sampling, compared with 17 using fast sampling, with eight using both modes.The review covered 98 papers, of which 75 engaged substantively in silicon sampling.
Motivation for fast sampling: Fast sampling uses fewer tokens and API calls to produce the same output tokens, making silicon samples cheaper and faster to assemble.Preprints commonly cite cost and scalability as reasons for adopting fast procedures.
Open question: Whether fast sampling produces samples as faithful as slow sampling remains an unresolved question with no known prior assessment.The literature identifies fidelity relative to slow sampling as the basis for fast sampling’s attractiveness.

3 Data and methods · 3.1 Data source

The study uses Singaporean World Values Survey data as human reference points for silicon samples, focusing on immigration attitudes and ethical norms. The measures use defined response scales, including reverse-coding for negatively phrased immigration items.

3.1 Data source: The study draws on two public-opinion batteries: eight immigration-effect items and 19 ethical-norm items.These batteries provide the opinion measures used to construct and assess silicon samples.
3.1.1 Key measures: Each immigration item uses a three-category response scale: disagree (0), hard to say (1), or agree (2).All eight immigration measures range from 0 to 2.
3.1.1 Key measures: Higher immigration-item values consistently indicate more favorable views of immigration.Four negatively phrased items are reverse-coded to maintain this direction.
3.1.1 Key measures: Ethical-norm items ask respondents whether each behavior or practice is justifiable using a ten-point scale, with higher values indicating greater acceptance.Summary statistics for the public-opinion items appear in Table 8.
3.1.1 Key measures: The reverse-coded immigration items concern increased crime, increased terrorism risk, increased unemployment, and social conflict.The four positively phrased items are left unchanged.
3.1.1 Key measures: Table 1 presents the costs of slow and fast silicon sampling.The supplied passage identifies the table but provides no numerical cost values.

3.2 Constructing silicon samples

The study constructs slow and fast silicon samples with GPT-5.4 by generating 30 surrogates per respondent from a nationally representative Singaporean survey sample. The approaches share initialization and output procedures but differ in whether survey items are queried repeatedly in segments or all at once.

Sample construction: 30 silicon surrogates were generated per respondent in each mode for 2,012 respondents, using GPT-5.4 with a 2,500-token maximum and temperature 1.0.The model weights were from March 5th, 2026.
Shared prompting procedure: Both modes use an initializing system and one-shot prompt, append the question prompt, and retrieve the model’s output.The initialization included survey context and an example output, with wording adapted from the WVS interviewer script.
Slow sampling: Slow sampling repeatedly submits segmented survey units and collects slices of the final output across non-cached API calls.The questionnaire is decomposed into atomic questions or small blocs appended to the initializing prompt.
Fast sampling: Fast sampling submits one elaborate prompt containing all survey items and returns the complete output in a single API call.Because each query produces a complete output set, responses are processed after every API query.

3.3 Estimands

The section defines three ecological estimands—population means, respondent-level Jaccard distances, and population variances—and specifies complementary measures for comparing opinion relationships and contextual spaces.

Estimands: The analysis compares population means, mean respondent-level Jaccard distances, and population variances across silicon samples.These are labeled A1, A2, and A3, respectively.
Population means: Mean bias is measured by comparing each silicon sample’s item-level sample mean, ¯x_k, with the corresponding WVS mean, ¯xWVS_k.Because immigration and morals use different scales, standardized bias divides each delta by the item’s WVS standard deviation, Δ_k/sWVS_k.
Respondent-level distance: Jaccard distance measures disagreement between each WVS respondent and their silicon surrogate, with D_i = 1 − J(t_i, ˜t_i) bounded from 0 to 1.The measure is averaged over 2,012 respondents and 30 resampling iterations within each domain.
Population variance: Population variance is assessed by comparing each attitude item’s silicon-sample standard deviation, s_k, with the corresponding WVS standard deviation using Fligner-Killeen tests.These tests require no normality assumption and are robust to heavily skewed, floor-bounded distributions.
Relationality and contextual space: Relational structure is compared using pairwise Cramer’s V associations and MCA contextual spaces aligned to WVS data with Procrustes analysis.Cramer’s V is used for categorical items; MCA represents opinion structure in a low-dimensional space, while Procrustes isolates shape differences after alignment.

4 Results · 4.1 How faithful are silicon estimates of population means? · 4.2 How faithful are silicon estimates of population variance?

Silicon samples produce moderately faithful estimates of population means, but substantially understate opinion variance. Fast sampling is generally comparable to, and sometimes modestly closer than, slow sampling on mean estimates, while both modes similarly poorly reproduce human variance.

4.1 How faithful are silicon estimates of population means?: Across 27 items, mean standardized bias was 0.40 SD for slow sampling and 0.37 SD for fast sampling, indicating moderate similarity to population means.The passage characterizes these estimates as typically moderately similar by conventional benchmarks.
4.1 How faithful are silicon estimates of population means?: Mean estimates were less faithful for immigration than moral attitudes: slow sampling showed 0.60 SD versus 0.31 SD, while fast sampling showed 0.47 SD versus 0.33 SD.Immigration attitudes carried roughly twice the standardized bias of moral attitudes under slow sampling and a smaller but still substantial excess under fast sampling.
4.1 How faithful are silicon estimates of population means?: Fast and slow sampling had similar mean errors: mean absolute delta was 0.605 versus 0.587, standardized bias was 0.397 SD versus 0.371 SD, and errors agreed in direction for 24 of 27 items (89%).The passage reports slow sampling first and fast sampling second.
4.2 How faithful are silicon estimates of population variance?: Silicon samples retained only 24% of real-world dispersion under slow sampling and 23% under fast sampling, with lower dispersion on every survey item.Both modes therefore substantially understate the variance of human opinions.
4.2 How faithful are silicon estimates of population variance?: Seven ethical-norm items and one immigration item collapsed to an SD of exactly 0.00 in both samples, reflecting degenerate homogeneity for those attitudes.The passage attributes this partly to LLMs’ tendency to collapse responses on serious moral transgressions.
4.2 How faithful are silicon estimates of population variance?: Slow and fast sampling had nearly identical aggregate retained variance, 23.9% versus 22.9%, while their item-level variance biases were strongly correlated at r= 0.984.Among non-degenerate items, fast sampling understated variance more often on ethical norms, while slow sampling produced the more homogeneous immigration sample; neither difference was large.

4.3 How faithful are bivariate associations? · 4.4 How faithful is the contextual space?

Silicon samples poorly reproduce both pairwise attitude associations and the broader contextual space underlying Singaporean opinions. Fast sampling is generally at least as faithful as slow sampling, while often achieving closer matches to WVS estimates.

4.3 How faithful are bivariate associations?: The study compares pairwise Cramer’s V matrices across slow and fast silicon samples, bootstrap WVS samples, and the WVS benchmark.Items without response variation are omitted because Cramer’s V is undefined.
4.3 How faithful are bivariate associations?: Immigration-attitude associations are poorly reproduced: silicon samples understate average co-variation, with mean deviation −0.167 for slow and −0.123 for fast.These deviations are measured relative to WVS estimates.
4.3 How faithful are bivariate associations?: Fast sampling better matches immigration associations, with correlation 0.42 versus 0.28 for slow and smaller average understatement, −0.123 versus −0.167.Both metrics favor fast sampling over slow sampling for immigration attitudes.
4.4 How faithful is the contextual space?: The broader contextual structure is assessed with two-dimensional MCA configurations separately fit to slow, fast, bootstrap, and WVS sources within each domain.Every attitude item is treated as an active categorical variable, and the first two dimensions are retained.
4.4 How faithful is the contextual space?: Because MCA spaces are invariant to translation, rotation, reflection, and scale, silicon configurations are Procrustes-aligned to WVS before measuring disagreement.The analysis reports Gower’s M2 and Tucker’s coefficient of congruence for each dimension.
4.4 How faithful is the contextual space?: Extreme homogeneity collapses 11 of 19 ethics items and 3 of 8 immigration items to a single category in both silicon samples.These items retain no residual variation for relating to other items; bivariate Cramer’s V does not reveal this because it is undefined for homogeneous variables.
4.4 How faithful is the contextual space?: Contextual spaces remain poorly matched to WVS, with Gower’s M2 of 0.610 immigration and 0.744 ethics for slow, versus 0.468 and 0.756 for fast.Fast is closer on immigration but not ethics under this dissimilarity measure.
4.4 How faithful is the contextual space?: Fast and slow contextual spaces are similarly different on ethics, while fast is closer on immigration and has higher Tucker’s φ on both immigration dimensions.Ethics: M2 is 0.744 slow versus 0.756 fast; immigration: φ is 0.769/0.688 fast versus 0.683/0.580 slow.

5 Conclusion

The conclusion presents silicon sampling as an early-stage method whose means may approximate public opinion, but whose distributions and relational structures remain insufficiently faithful. It also identifies fast sampling as conditionally superior to slow sampling and geometric analysis as a useful complement to conventional evaluation.

Conclusion: Silicon sampling remains an innovative method in early development and cannot faithfully reproduce Singaporean respondents’ answers across immigration and ethical-norm questions.Its estimates of mean public opinions can nevertheless appear surprisingly good.
Conclusion: Fast silicon sampling produces samples no worse in fidelity than conventional slow sampling methods, supporting its conditional superiority.The conclusion identifies this as the first of three important implications for silicon-sampling methodology.
Conclusion: Silicon samples should be treated as exploratory instruments at best, not substitutes for human respondents in substantive claims about Singaporeans.The conclusion highlights limited fidelity despite reassuringly close means, particularly regarding opinion dispersion and relational structure.
Conclusion: Geometric analysis supplements model evaluation by measuring divergences in relational fidelity that item-level means and pairwise associations can leave invisible.A silicon sample may approximate many margins while assembling them into an opinion space no human population would recognize.

6 Appendix I: Descriptive Statistics

Appendix I reports descriptive statistics for key demographic covariates and World Values Survey (WVS) attitude variables.

Table 7 provides summary statistics for key demographic covariates.
Table 8 provides summary statistics for WVS attitude variables.

7 Appendix II: Cramer’s V

Appendix II presents Cramer’s V analyses of pooled immigration and moral attitude items, including panels for fast silicon sampling and the World Values Survey.

Cramer’s V: Table 9 reports Cramer’s V between immigration attitude items across all iterations pooled.
World Values Survey (WVS): The appendix identifies a World Values Survey (WVS) panel.
Cramer’s V: Table 10 reports Cramer’s V between moral attitude items across all iterations pooled.
Fast silicon sampling: The appendix identifies a Fast silicon sampling panel.
World Values Survey (WVS): A second panel is identified as the World Values Survey (WVS).

8 Appendix III: MCA Contributions

Appendix III reports multiple-correspondence-analysis category contributions for immigration and moral attitude items, including comparisons involving fast silicon sampling and the World Values Survey.

Immigration attitude items: Table 11 reports MCA category contributions for immigration attitude items.
Reported fields: The reported MCA fields include Item, Cat., Coord., Ctr (%), and cos2.
Moral attitude items: Table 12 reports MCA category contributions for moral attitude items.
Panels: The appendix includes panels for Fast silicon sampling and the World Values Survey (WVS).