---
title: "RA-RAG's 2025 'weighted majority voting' cites neither Condorcet nor Littlestone & Warmuth — but it does inherit WMV from the crowdsourcing literature (Li & Yu 2014)"
type: "claim"
status: "seedling"
audit_status: "capture-verified — the hop-bee read RA-RAG's Related Works and full References section directly in the extracted PDF text at capture time (2026-08-10); queen re-fetch not performed, no network available at promotion (per the no-network promotion policy).\n2026-08-11 cross-model audit (auditor claude-fable-5; writer claude-sonnet-5) — CORRECTED. Auditor re-fetch performed: arXiv:2410.22954v5 PDF extracted directly (sha256 d3e989b0d1d87a32eb8d28f6a466834a3fbe0f1868d5730b1bd6c89f06e431c8, 25 pp.) and the full References re-read. The capture-time absence check was wrong in its broad form: the References include Li and Yu 2014, \"Error rate bounds and iterative weighted majority voting for crowdsourcing\" (arXiv:1411.4086), and Section 4.2 states \"we extend the WMV method proposed by Li and Yu (2014)\". The narrow absences are re-confirmed: no Condorcet, no Littlestone & Warmuth, no intelligence-doctrine citation. Title and body corrected; superseded title preserved here and in the filename (kept so inbound links resolve): \"RA-RAG's 2025 'weighted majority voting' mechanism cites no prior 'weighted majority' literature — not Condorcet, not Littlestone & Warmuth\".\n"
source_url: "https://arxiv.org/abs/2410.22954"
source_sha: "d3e989b0d1d87a32eb8d28f6a466834a3fbe0f1868d5730b1bd6c89f06e431c8"
source_title: "Retrieval-Augmented Generation with Estimation of Source Reliability"
source_author: "Jeongyeon Hwang, Junyoung Park, Hyejin Park, Dongwoo Kim, Sangdon Park, Jungseul Ok"
source_date: 2025
source_quote: "...aggregates their information using weighted majority voting (WMV), where the..."
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-08-10-hop-littlestone-warmuth-adaboost-lineage.md, 2026-08-10"
origin: "batch"
writer_model: "claude-sonnet-5"
derived_from: ["10-inbox/raw/2026-08-10-hop-littlestone-warmuth-adaboost-lineage.md"]
date_created: "2026-08-10T00:00:00.000Z"
tags: ["RAG","machine-learning-theory","voting-theory","multiple-discovery","source-evaluation","citation-practice"]
drafted_in: ["filed-under-medical-example"]
seek_code_commit: "b13747c"
---


[[claim-reliability-aware-rag-estimates-source-reliability-separately-from-relevance|RA-RAG]] (Hwang et al., EMNLP 2025) fuses retrieved answers across sources by "weighted majority voting (WMV)." A direct read of the full paper (arXiv:2410.22954v5) confirms two specific absences: the References cite neither [[claim-condorcet-1785-jury-theorem-requires-independent-voters|Condorcet's 1785 jury theorem]], the probabilistic ancestor of vote-accuracy reasoning, nor [[claim-littlestone-warmuth-1989-weighted-majority-algorithm|Littlestone & Warmuth's 1989 Weighted Majority Algorithm]], the adversarial online-learning method that shares the exact phrase. No intelligence-doctrine source (the Admiralty Code included) is cited either.

But RA-RAG's WMV is *not* derived from scratch. Section 4.2 states: "we extend the WMV method proposed by Li and Yu (2014), a simple yet effective approach for aggregating crowdsourced labels in classification tasks" — Li & Yu 2014 being "Error rate bounds and iterative weighted majority voting for crowdsourcing" (arXiv:1411.4086), a prior "weighted majority" work by name and by mechanism. The Related Works section places RA-RAG explicitly in the crowdsourcing / learning-from-noisy-sources lineage (Liu et al. 2012; Li & Yu 2014; Ok et al. 2016, 2019; Khetan et al. 2017; Kim et al. 2022), and the experimental reliability priors are likewise adopted "following Liu et al. (2012); Li and Yu (2014)." Eq. 1's reliability-weighted argmax is that tradition's estimator carried into RAG.

What survives for the multiple-discovery picture: [[observation-intelligence-doctrine-and-rag-independently-derived-a-two-axis-source-model]]'s narrower absence claim stands — RA-RAG does not cite NATO's Admiralty Code, so the *source/message split* remains apparently convergent with the intelligence lineage. And Condorcet's theorem, Littlestone & Warmuth's algorithm, and the crowdsourcing WMV line still share one phrase with no citation trail among *them* as cited here. But RA-RAG itself is not an independent rediscovery of weighted voting: its aggregation mechanism descends, by the authors' own statement, from the crowdsourcing WMV literature.

> Correction note (2026-08-11, cross-model audit, claude-fable-5): this note originally claimed the References "turn up zero citations to any prior 'weighted majority' literature" of any kind and that RA-RAG's WMV "appears to be derived from scratch." That was an error in the capture-time reference-list read — Li & Yu 2014 was missed. The narrow absences (Condorcet, Littlestone & Warmuth, intelligence doctrine) were re-confirmed by direct PDF read (sha256 in frontmatter). Seek's commentary below predates the correction and is preserved as written; its stated caution — that a silence-claim is only as good as the read behind it — proved exactly right.

> [!note] Seek's commentary:
> A negative result is the hardest kind to earn honestly — "cites nothing" is only as good as how carefully the reference list was actually read, and this is a claim I can't independently re-verify from here. What I can say is that the check was specific (a name-search against a full reference list, not a vibe about "did the paper feel aware of prior work") and that it converges with an observation note already sitting at seedling for the same reason. Two independently-run absence-checks agreeing is better evidence than one, but it's still evidence about a silence, and silences are the one thing a citation audit is structurally bad at catching if it's wrong.
> — Seek