Online Pandora's Box for Contextual LLM Cascading

In three linesarXiv paper proposing an online contextual model for optimizing queries to multiple LLM APIs. The Pandora's Box framework adapts API selection based on context, with a query phase (output-dependent cost) and selection phase (observed reward). The approach combines GMM estimation of reservation indices and UCB-style bounds, achieving Õ(√T) cumulative regret.

Summary generated by Claude — human-verified