A recommendation method reported higher offline ranking scores than the strongest baseline in 41 of 48 corrected settings, or 85.42%, while adding far less measured request-time overhead than REARANK. Called CRAMER, it is designed to let a natural-language request steer a sequential recommender while the underlying model stays frozen.
The method targets a practical tension in recommender systems: preserving a trained model's existing behavior while responding to an immediate request. The paper frames the task as controlling a frozen sequential-recommender backbone rather than rebuilding it for every new instruction.
A request layer over a fixed model
CRAMER maps request text to lightweight binary gate vectors, or on/off controls, that become masks over the frozen backbone. The backbone itself remains unchanged during training; learning is limited to a projection layer and a selected subset of request-encoder parameters.
The training objective includes an auxiliary sparsity prior. The authors treat the resulting loss as a variationally inspired surrogate rather than a strict ELBO, or evidence lower bound.
How the offline test was built
The evaluation drew on four processed datasets: ReDial, with 5,207 items and 36,460 interactions; KuaiSAR, with 174,895 items and 260,243 interactions; Beauty, with 44,977 items and 122,485 interactions; and CDs&Vinyl, with 76,368 items and 141,213 interactions. The authors built request text from prior interactions, kept positive interactions or ratings of at least 4.0 depending on the dataset, and randomly downsampled three larger datasets because of computing limits.
For each offline test, one ground-truth positive was paired with 100 randomly sampled items the user had not interacted with. Hit rate, NDCG and mean reciprocal rank were reported at cutoffs of 10 and 20, asking in different ways whether the known item appeared near the top and how high it ranked. That protocol makes the headline result a sampled-candidate ranking comparison, not a test of live user behavior.
Scores and speed in the benchmark
In a representative SASRec results block, CRAMER's hit-rate score at 10 was 0.578, compared with 0.549 for REARANK. At a cutoff of 20, the scores were 0.694 and 0.684. The displayed values were averaged over five runs.
Across 48 paired t-tests, the researchers applied Benjamini–Hochberg correction to control the false-discovery rate at 0.05. After that adjustment, CRAMER was significantly better than the strongest baseline in 41 settings, or 85.42% of the total.
The efficiency comparison favored the lighter control layer. With SASRec, CRAMER added 0.018 seconds and 1,355.6 MiB of GPU memory per request, while REARANK added 9.256 seconds and 9,824.7 MiB. The vanilla SASRec figures were 0.033 seconds and 2,024.1 MiB.
The same pattern appeared with BERT4Rec. CRAMER added 0.021 seconds and 1,408.1 MiB per request, against 9.184 seconds and 9,412.5 MiB for REARANK; the vanilla backbone measured 0.038 seconds and 2,119.6 MiB.
Requests changed the model's output mix
The study also tested whether requests could push the model's output in a recognizable direction. In an analysis of 100 users, romance-related items made up 0.286 of the top-10 recommendations with no request, 0.432 after a positive romance request and 0.135 after an avoidance request. The figures show request-conditioned movement in the recommendation mix, not user satisfaction.
In a controlled scaling study on KuaiSAR, increasing BERT4Rec depth from 4 to 10 layers was accompanied by a rise in NDCG@10 from 0.366 to 0.462. The reported runtime ratio fell from 55.3% to 41.5%, and the memory ratio from 66.4% to 43.7%.
A separate case study of five selected CDs&Vinyl users found CRAMER at the highest rank for all of them under both tested backbones. That result is an illustration of per-user alignment, but its small, selected sample cannot stand in for broad performance.
A promising result with a narrow test
The evidence covers four processed datasets, frozen Transformer-based sequential recommenders and a sampled-negative evaluation design. It does not show improved human satisfaction, engagement or business outcomes, and it does not establish how the masks would behave under other request wording, domains or architectures.
Paper data and sources
Original title: CRAMER: Control via Request-Aware Masking for Editing Recommenders
Authors: Zhiyuan Julian Su, Naihe Feng, Zhen Luther Qin, Ga Wu
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text