arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2601.00841cs.LG
收藏

基于SLO的检索增强生成中的动作路由:目标消融与失败模式

SLO-Conditioned Action Routing for Retrieval-Augmented Generation: Objective Ablation and Failure Modes

Medium

Bharath Nunepalli

AI总结:

本研究提出基于SLO的RAG控制方法,通过动作路由优化检索深度和生成模式,在质量导向SLO下实现成本节约,但在廉价SLO下可能导致拒绝崩溃。

摘要:

检索增强生成(RAG)引入了一个实际的控制问题:检索深度和生成行为必须根据查询进行选择,以满足服务等级目标(SLOs)如成本、拒绝率和幻觉风险。本工作将每查询控制建模为一个小的离散动作:选择检索深度和生成模式(受控 vs. 自动),或拒绝。通过执行每个动作并记录准确性、令牌成本、幻觉/拒绝指标以及SLO加权奖励,构建了一个离线日志数据集。评估了两个简单的策略学习目标:对每个状态最佳动作的监督分类(Argmax-CE)和一个奖励加权变体(Argmax-CE-WT)。在评估的设置中,一个强固定基线(低k,受控提示)表现竞争;学习策略主要在质量导向的SLO下提供额外的成本节约,但在廉价SLO下当拒绝被大量奖励时会出现拒绝崩溃。贡献是RAG流水线中SLO意识控制的可重复案例研究,强调失败模式和报告惯例,而不是提出新的检索器或语言模型。

英文摘要:

Retrieval-augmented generation (RAG) introduces a practical control problem: retrieval depth and generation behavior must be chosen per query to satisfy service-level objectives (SLOs) such as cost, refusal rate, and hallucination risk. This work models per-query control as a small discrete action: choose a retrieval depth and a generation mode (guarded vs. auto), or refuse. An offline logged dataset is constructed from SQuAD 2.0 by executing each action and recording accuracy, token cost, hallucination/refusal indicators, and an SLO-weighted reward. Two simple policy-learning objectives are evaluated: supervised classification of the per-state best action (Argmax-CE) and a reward-weighted variant (Argmax-CE-WT). Across the evaluated settings, a strong fixed baseline (low k, guarded prompting) performs competitively; learned policies mainly provide additional cost savings under a quality-focused SLO and can exhibit refusal collapse under a cheap SLO when refusal is heavily rewarded. The contribution is a reproducible case study of SLO-aware control for RAG pipelines, emphasizing failure modes and reporting conventions rather than proposing a new retriever or language model.

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。

↑