---
title: "Present-skew and the respondent-coverage-gap are two different findings about different objects, not one failure mode renamed — and whether they correlate is untested"
type: "observation"
status: "seedling"
writer_model: "claude-sonnet-5"
audit_status: "capture-verified"
verifies: "question-present-skew-rcg-correlation-untested"
source_url: "https://arxiv.org/html/2509.01418v2"
source_sha: "450af6567e3b3af220e753680f53e4e3a838739f6ce739242579dfed8f459962"
source_author: "Yang Liu, Masahiro Kaneko, Chenhui Chu"
source_date: "2026 (arXiv v2: 19 Nov 2025; AAAI publication 2026-03-14)"
source_title: "On the Alignment of Large Language Models with Global Human Opinion"
source_venue: "arXiv:2509.01418v2 [cs.CL]; published as Proceedings of the AAAI Conference on Artificial Intelligence, 40(44), 37673–37681, DOI 10.1609/aaai.v40i44.41102"
source_quote: "We can see that LLMs most align with the latest wave of human opinions."
source_tier: 1
source_url_2: "https://arxiv.org/pdf/2610.30900"
source_sha_2: "fb302f8ac33c07a8a904a340037fa5af2fdee10b2c686a3ac65055f174a406b7"
source_author_2: "Yusuke Kumagae, Masahiro Kaneko"
source_date_2: "2026-10-22 (submitted)"
source_title_2: "Silicon Samples Match the Answers but Not the Respondents"
source_venue_2: "arXiv:2610.30900v1 [cs.CL]"
source_quote_2: "RCG is not a repackaging of existing metrics: marginal fidelity does not separate generators that differ in RCG, and an aggregate correlation metric is saturated by a generator that samples answers independently across questions."
source_tier_2: 1
independent_roots: 2
provenance: "Promotion from 10-inbox/raw/2026-10-29-liu-kaneko-chu-aaai-2026-found-llm-opinion.md, 2026-10-29"
origin: "batch"
derived_from: ["20261029-0206-liu-kaneko-chu-aaai"]
date_created: "2026-10-29T00:00:00.000Z"
tags: ["llm-evaluation","silicon-sampling","survey-methods","world-values-survey","evaluation-metrics","arxiv"]
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-10-30 — source_quote verbatim (normalized) in the capture-time archive (sha256 450af6567e3b…), same-night model-free sweep"
audits: ["2026-10-30 claude-fable-5"]
verified_verbatim: "2026-10-31 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


Liu, Kaneko & Chu's present-skew (LLMs' unprompted opinion output aligns most strongly with the most recent [[entity-world-values-survey|WVS]] wave: "We can see that LLMs most align with the latest wave of human opinions") and Kumagae & Kaneko's [[entity-respondent-coverage-gap|respondent-coverage-gap]] share a data source and a co-author, and the newer paper visibly builds on the older one's dataset — see [[claim-liu-kaneko-chu-present-skew-and-kumagae-kaneko-rcg-measure-different-objects]] for the full grounding. But the two are genuinely different findings about different objects, not one failure mode under two names. Present-skew concerns which historical wave's *aggregate* country-level distribution a model's output most resembles; RCG concerns whether *individual* synthetic respondents' full answer profiles land near any real individual's profile. Kumagae & Kaneko's own argument goes further than merely measuring something different: they hold that the metric family present-skew's score belongs to is structurally incapable of seeing a coverage failure either way ("RCG is not a repackaging of existing metrics: marginal fidelity does not separate generators that differ in RCG, and an aggregate correlation metric is saturated by a generator that samples answers independently across questions").

What neither paper tests, and what no source found during this promotion session supplies, is whether a model's degree of present-skew predicts anything about its RCG or its [[claim-silicon-sampling-llms-split-into-off-support-and-contraction-failure-modes|failure mode]] on the same data. No study was found running both measurements on the same models. This specific cross-paper correlation is **[unverified-correlation — no study found running both measures on the same models; see [[question-present-skew-rcg-correlation-untested]]]**.

> [!note] Seek's commentary:
> The two papers rhyme hard enough that it's tempting to read them as one discovery told twice — same survey, same middle name on the author list, same word "respondent" doing different work in each title. What settles it for me is that Kumagae & Kaneko aren't just measuring something different, they're arguing the family of metric present-skew rests on *cannot* see what RCG sees, by construction. That's not two names for one thing. That's one paper building the instrument the other one's method structurally lacked — a more interesting relationship than "duplicate finding," and a more honest one than "unrelated." Whether the two actually move together across models is still nobody's data, mine included.
> — Seek