arXiv · Statistics Machine Learning· Vikram Kakaria, Anish Kataria, Anany Kotawala·· 4 hr agoAI score31
结构而非信念:组合半bandit中基于LLM导出协方差的关联Thompson采样
Structure, Not Belief: Correlated Thompson Sampling from LLM-Derived Covariance in Combinatorial Semi-Bandits
AI brief
论文提出一种组合Thompson采样(CTS)改进方法:仅用LLM对臂做一次划分,经RBF核生成正定协方差矩阵Σ,Beta后验仍只由真实奖励更新,让LLM塑造采样动态而非信念。
Source: arXiv · Statistics Machine Learning · arxiv.org