Challenges and Solutions for Bandits in the Wild: Warm-Started Mixture Bandits for Cross-Cohort Slate Recommendation¶
L1 core-mechanism diagnostic. Implements fixed-strength mixture priors, Thompson slates, and posterior updates, not historical matrix factorization or the 25-day deployment.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2609.37800 |
| Organization | DFKI / RPTU Kaiserslautern-Landau |
| First public release | 2026-09-29 (arXiv v1) |
| Authors' code | Official public code |
| Adapter | cohortmix-ts |
| Local reproduction code | src/auto_research/reproductions/cohortmix_ts/ |
Original paper¶
Motivation and method¶
Learn user groups and item arms from historical cohorts. Use metadata to build a fixed-strength mixture Beta prior for a new user. At each round, Thompson sampling selects a previously unexposed slate; feedback updates each user's posterior independently.
flowchart LR
I[Public input and current state] --> M[CohortMix-TS mechanism]
M --> A[Auditable intermediates]
A --> O[Output or state update]Key figure from the paper¶
Figure 3: Original method design and components. From the paper; copyright remains with the authors.
Core equation¶
\(\hat\mu_{c,a}=(\alpha_0+s_{c,a})/(\alpha_0+\beta_0+s_{c,a}+f_{c,a})\). The new user's prior uses fixed-strength Beta pseudo-counts weighted by cohort membership.
Paper results¶
In the 25-day randomized deployment, the between-group difference in early-to-late correctness change was 6.23 percentage points in the restricted full-window subgroup, with 95% bootstrap interval [0.5, 11.9] and p=0.043. Engagement among all registered users did not differ significantly.
Local reproduction¶
Local comparison: The baseline disables the mechanism or uses the default state, and the treatment enables the core operator. This checks invariants and state transitions only; cross-model relative changes are not applicable.
See the three-seed diagnostic. diagnostic_only=true indicates executable state transitions, gradients, or scheduling invariants only, not eligibility for capability ranking.
Reproduction boundaries¶
Fixed-strength mixture priors, Thompson slates, and posterior updates are implemented. Historical matrix factorization and the 25-day deployment were not reproduced.