Evaluate richer transition and replay models for level and contrast dynamics

Evaluate level-transition models with partial pooling, publicly available traffic-log replays containing platform-wide changes, and dynamic models for the contrast vector that preserve the distinction between arm comparisons and the common level.

Background

The basic Odds-Ratio Thompson Sampling specification treats the batch-level intercept as entirely fresh and the contrasts as persistent, optionally with scalar decay. The paper proposes richer alternatives, including structured or partially pooled level transitions and dynamic contrast models, but does not evaluate them comprehensively. It also identifies a need for replay studies using public traffic logs that capture genuine platform-wide changes, where the common-level component is especially relevant.

References

Also open, and next in line: racing the level transitions of Section~6.2, partial pooling included, across batch sizes, a registered replay using publicly available traffic logs that capture a platform-wide change, and a dynamic model for $_t$ that keeps the OR-TS distinction between comparison and level.

— Odds-Ratio Thompson Sampling: A Specification and Design Guide for Contrast-Based Multi-Armed Bandits  (2609.19709 - Kim, 17 Sep 2026) in Discussion, final paragraph (“Five bounds remain”)