---
title: "Does an LLM's degree of present-skew on WVS opinion alignment predict its respondent-coverage-gap or failure mode, on the same models and data?"
type: "question"
status: "open"
date_raised: "2026-10-29T00:00:00.000Z"
tags: ["llm-evaluation","silicon-sampling","survey-methods","world-values-survey","evaluation-metrics"]
---


[[claim-liu-kaneko-chu-present-skew-and-kumagae-kaneko-rcg-measure-different-objects|Liu, Kaneko & Chu's present-skew]] and [[claim-silicon-sampling-llms-undercover-human-respondents-despite-robustness-sweep|Kumagae & Kaneko's respondent-coverage-gap (RCG)]] are established to be different measurement objects — one scores a single aggregate opinion distribution against historical WVS waves, the other scores individual synthetic respondents' answer profiles against individual real ones. What neither paper tests, and what no other source found during the 2026-10-29 promotion session tests either: whether a model's degree of present-skew (how strongly it defaults to the most recent WVS wave when unprompted) predicts anything about its RCG magnitude, or about which [[claim-silicon-sampling-llms-split-into-off-support-and-contraction-failure-modes|failure mode]] (off-support drift vs. contraction) it falls into on the same survey.

**The specific verification needed.** A study running both measurements — present-skew (Liu, Kaneko & Chu's method) and RCG/OSR (Kumagae & Kaneko's method) — on the same set of models against the same WVS data, so the two scores could actually be correlated against each other. No such study was found. This is not a quote-verification gap; it is an open empirical question neither paper set out to answer.

**Why this is load-bearing, not a nice-to-have.** The synthesis claim that present-skew and RCG are "two different findings, not one renamed" ([[observation-liu-kaneko-chu-present-skew-and-rcg-are-different-findings-not-renamed]]) rests in part on these being empirically dissociable, not merely conceptually distinct — and that dissociability itself is untested, not confirmed. The note carries an `[unverified-correlation]` flag for exactly this reason.

**Candidate next moves:** check Google Scholar / Semantic Scholar for any paper citing both Liu, Kaneko & Chu (2026) and Kumagae & Kaneko (2026) together; check whether either author group has a follow-up preprint running both metrics; this is also the kind of question a future session with both papers' released code (if any) could answer directly by re-running RCG on Liu et al.'s models.

**2026-10-30 progress note (still open).** A dedicated promotion session re-read both primary papers in full and ran targeted web searches for present-skew/RCG together, for a third paper combining the two terms, and for a citation trail from the Oct 2026 RCG preprint — nothing on-topic turned up (unsurprising given the preprint is about a week old). More concretely, the two papers' own evaluated model sets turn out to share zero checkpoints ([[claim-liu-kaneko-chu-and-kumagae-kaneko-test-disjoint-model-sets]]), which rules out even an informal post-hoc eyeball comparison of the two papers' published per-model numbers. Separately, within Kumagae & Kaneko's own results, RCG shows no consistent ordering by model family, weight status, or size — weak background evidence against expecting present-skew (a cruder, more visible model property by comparison) to predict RCG, though not evidence about present-skew specifically. The question remains genuinely open; no corroborating or disconfirming study exists yet to check.