Fast A/B/n Testing: Exact Multi-Policy Comparison via Tree-Coupled Feedback Sharing¶

Small NumPy mechanism diagnostic, not industrial A/B/n or full benchmark.

Paper information¶

Field	Details
Paper	Paper or original source
Institution	New York University; first author Yuxiao Wen.
First public release	2026-08-13 (arXiv v1)
Author code	No public author code found as of 2026-08-24.
Local adapter or method key	`tcab`
Local reproduction code	`src/auto_research/reproductions/tcab/`

Paper method and evidence¶

Use maximal coupling and a minimum spanning tree to share feedback for matching actions while preserving each adaptive policy's marginal trajectory distribution.

flowchart LR
  A[Adaptive policies] --> B[Maximal coupling and tree]
  B[Maximal coupling and tree] --> C[Share feedback for same decisions]

Original paper figure¶

Figure illustrates coupled feedback sharing. From the original paper; copyright remains with the authors.

Core equation¶

\[ N(T)=T+\sum_{t,e}D_{e,t} \]

Reported results¶

Theory reduces independent A/B/n feedback from JT to T+o(T); experiments improve the cost–precision frontier. Not local online evidence.

Local reproduction¶

Same 64 examples and seed: independent A/B/n feedback accuracy 1.0000, TCAB 1.0000, +0.00 percentage points.

auto-research reproduce --paper tcab --seed 42

Fixed metrics. Batch index. No paper-scale checkpoint, custom CUDA kernel or full benchmark.