towerbench: A frozen-protocol benchmark harness for two-tower retrieval and autoencoder recommenders with paired bootstrap statistics on public implicit-feedback data

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

We present towerbench , an open-source Python/PyTorch harness that benchmarks two-tower retrieval models against linear and variational autoencoders, sequential and graph recommenders under a frozen evaluation protocol. Every experiment ranks the full catalogue, the test split is read exactly once and logged, and model differences are reported as paired per-user bootstrap intervals across seeds and rolling-origin temporal windows rather than as single point estimates. Each industrial two-tower technique (logQ-corrected in-batch softmax, mixed negative sampling, MixGCF hard negatives, DCN-v2 cross layers, frozen sentence-transformer item text, geo/price side features) is one configuration switch. Seven datasets from five public sources (MovieLens, Amazon Reviews 2023, Gowalla, Inside Airbnb, Steam) ship with reproducible loaders, a full sweep of 1974 experiments runs unattended on two GPUs in about a day, and a configuration-sensitivity check shows which of its verdicts survive retuning.