何时委托优于多数？一种基于委托的多样本LLM推理聚合器
When Does Delegation Beat Majority? A Delegation-Based Aggregator for Multi-Sample LLM Inference
AI总结：
提出基于委托的聚合器PPV，利用样本的字母熵和推理几何信号，在MMLU-Pro上比多数投票高1.5个百分点，无需标签或训练。
摘要：
多数投票是对多样本LLM推理进行无监督聚合的主流方法。我们证明，将每个样本携带的信号输入基于委托的聚合器（传播代理投票，PPV）可产生一种无监督共识规则，在MMLU-Pro上整体比多数投票高1.5个百分点，在非平凡子集上高2.24个百分点（配对McNemar p ~ 1.0e-14，n = 8,099）。多数投票丢弃了每个样本携带的两个自由信号：组内字母熵和组间推理几何。PPV暴露了两个每个投票者使用的杠杆，它们恰好消耗这些信号：WHEN（投票者保留自己选择的权重）和WHOM（如何将剩余权重分配给同行）。我们使用字母熵驱动WHEN，使用以问题为中心的嵌入余弦驱动WHOM。该方法不需要真实标签和辅助训练：对于每个问题，我们将128个采样生成划分为16组，计算每组的字母级语义熵和推理嵌入质心，并将两者输入随机委托矩阵，其平稳分布选择共识答案。我们通过一个例子说明PPV如何推翻一个明显的10-6多数（错误答案）：10票的多数簇几何上不连贯（平均簇内余弦-0.02），而6票的少数簇紧凑（+0.26），因此传播的委托质量集中在少数派的答案上，尽管仅凭熵会使多数保持领先。我们还报告了具有负面结果的委托策略，这些策略限制了无监督LLM聚合的设计空间：没有问题内的置信度模式集成能够缩小与oracle的差距。
英文摘要：
Majority voting is the default unsupervised aggregator for multi-sample LLM inference, but it discards two signals: within-group answer entropy and between-group reasoning geometry. We aggregate by delegation instead (Propagational Proxy Voting, PPV): each group of samples keeps weight on its own answer in proportion to its entropy-based confidence (When) and routes the rest to peers by reasoning-embedding similarity (Whom); the stationary distribution of the resulting delegation matrix picks the consensus answer. This requires neither gold labels nor training. On MMLU-Pro with 128 samples per question, delegation beats majority by +1.5 pp overall and +2.24 pp on non-trivial questions (McNemar p ~ 1.0e-14, n = 8,099), overturning wrong majorities whose answer cluster is geometrically incoherent while the correct minority is tight. We then characterize exactly when delegation overturns majority: a two-option model gives a closed-form flip condition on each option's confidence and the weight it routes to the other, with a do-no-harm corollary for near-unanimous questions. The condition calls the realized winner on 96.5% of non-trivial questions, and its predicted mass gap tracks the realized gap at r = 0.97. We did not find any other unsupervised ensemble methods that close the oracle gap.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
