Abstract:Majority voting over sampled answers is the dominant unsupervised aggregator for multi-sample LLM inference. We show that piping the signals every sample carries into a delegation-based aggregator (Propagational Proxy Voting, PPV) yields an unsupervised consensus rule that beats majority on MMLU-Pro by +1.5 pp overall and +2.24 pp on the non-trivial subset (paired McNemar p ~ 1.0e-14, n = 8,099). Majority discards two free signals every sample carries: within-group letter entropy and between-group reasoning geometry. PPV exposes two per-voter levers that consume exactly these signals: WHEN (how much weight a voter keeps on its own pick) and WHOM (how it splits the remainder across peers). We drive WHEN with letter entropy and WHOM with per-question-centered embedding cosine. The method needs no gold labels and no auxiliary training: per question, we partition 128 sampled generations into 16 groups, compute each group's letter-level semantic entropy and reasoning embedding centroid, and feed both into a stochastic delegation matrix whose stationary distribution selects the consensus answer. We walk through an example in which PPV overturns a clear 10-6 majority for the wrong letter: the 10-voter majority cluster is geometrically incoherent (mean within-cluster cosine -0.02) while the 6-voter minority is tight (+0.26), so propagated delegation mass concentrates on the minority's answer even though entropy alone would keep the majority ahead. We further report delegation strategies with negative results that constrain the design space for unsupervised LLM aggregation: no within-question ensemble of confidence modes closes the oracle gap.
From: Yasushi Sakai [view email]
[v1]
Sat, 6 Jun 2026 10:57:45 UTC (106 KB)
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。