Where Do LLM Ensemble Gains Come From? Decomposing Candidate Value, Diversity, and Aggregation
The paper says ensemble gains depend on the control condition used to measure them.
Donghwan Kim compares an LLM aggregator across direct answering, a single candidate, a duplicated candidate, and separately generated candidates. The study finds that the same overall candidate gain can split differently into diversity and context effects depending on whether the baseline is single-slot or duplicate. On MultiHopRAG and FRAMES, the “repetition gap” even changes sign in the stronger-panel setting. The stated takeaway is procedural: ensemble evaluations should name their reference control before assigning gains to extra candidates. ArXiv · AI/CL/LG's note
Donghwan Kim compares an LLM aggregator across direct answering, a single candidate, a duplicated candidate, and separately generated candidates. The study finds that the same overall candidate gain can split differently into diversity and context effects depending on whether the baseline is single-slot or duplicate. On MultiHopRAG and FRAMES, the “repetition gap” even changes sign in the stronger-panel setting. The stated takeaway is procedural: ensemble evaluations should name their reference control before assigning gains to extra candidates. ArXiv · AI/CL/LG's note
score 4