跳到正文
HuggingFace Daily Papers(社區熱門論文)·· 16 天前AI 評分39

PISA 與 PISA-E:面向提示詞自適應、成本感知 LLM 推理的集成路由

Ensemble Routing for Prompt-Adaptive Cost-Aware LLM Inference

AI 導讀

這項研究提出單模型路由器 PISA,以及可按提示詞調整組成和規模的集成路由器 PISA-E,後者毋須專門的集成訓練。兩者在四個路由基準的分佈內及分佈外設定下接受評估;PISA-E 在優先考慮品質時,集成任務表現的宏平均最高。相較最強的競爭集成方案,PISA-E 亦提供平均推理成本約低 36%(ID)及 25%(OOD)的運作選項。

正文

Most LLM routing approaches select a single model per prompt, while ensembling introduces a different question: which models should we combine, and how many are actually worth invoking?

In this paper, we introduce PISA, a cost-aware single-model router, and PISA-E, which extends routing to ensembles with prompt-adaptive composition and size.

PISA-E estimates ensemble utility from individual model capability predictions, without requiring ensemble-specific training. It then selects the models to invoke based on the prompt and a configurable cost–quality preference.

We evaluate both methods across four routing benchmarks under in-distribution and out-of-distribution settings. PISA achieves the highest macro-average cost–performance hypervolume among the evaluated single-model routers. PISA-E achieves the highest macro-average ensemble task performance when prioritizing quality, while also offering operating points with higher performance at approximately 36% (ID) and 25% (OOD) lower average inference cost than the strongest competing ensembles.

Feedback and discussion are welcome, particularly on the ensemble-selection formulation and adaptive ensemble cardinality.

來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co