arXiv：cs.AI· Donghwan Kim·· 20 小时前
LLM 集成收益从何而来？拆解候选价值、多样性与聚合
Where Do LLM Ensemble Gains Come From? Decomposing Candidate Value, Diversity, and Aggregation
AI 导读
研究在 Direct、Single-slot、Duplicate、Multi-candidate 四种匹配条件下考察 LLM 聚合器，发现整体候选收益 CIV 可代数拆分为额外候选对比 DSV 与候选上下文对比 CCV，且该拆分（而非 CIV）取决于所选对照条件。
正文
Abstract:LLM ensemble gains are often measured against a single baseline, but the value attributed to additional candidates can change with the comparison used. We study an LLM aggregator under four matched conditions: Direct (no candidates), Single-slot (one candidate in a multi-slot interface), Duplicate (one candidate repeated), and Multi-candidate (separately generated candidates). Under either non-diverse control, overall candidate gain (CIV) splits algebraically into an additional-candidate contrast (DSV) and a candidate-context contrast (CCV); the split, but not CIV, depends on the chosen control. On MultiHopRAG model diversity (k=2), the repetition gap Rep=-8.33 pp [-13.25, -3.75] quantifies the difference between gains measured against Duplicate and Single-slot. On a matched FRAMES stronger-panel extension (n=200, k=3, aggregator output limit 4096), paired Delta Rep=+9.67 pp [+6.00, +13.50]; within panels, Rep changes sign (original -6.33; stronger +3.33). In FRAMES evidence aggregation, candidate-context contrasts remain strongly negative under both non-diverse controls. Our results show why ensemble evaluations should state the reference control before attributing gains to additional candidates.
| Comments: | 14 pages, 6 figures |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.18786 [cs.CL] |
| (or arXiv:2610.18786v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.18786 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Donghwan Kim [view email]
[v1]
Tue, 13 Oct 2026 17:36:51 UTC (104 KB)
来源：arXiv：cs.AI · arxiv.org