A model that listens to the request
CRAMER is designed to adapt a frozen sequential recommender, a system that uses a user's previous interactions to rank items, to natural-language requests through parameter masking. In the reported comparisons, it was the best-performing method on all metrics and was significantly better than the strongest baseline in 41 of 48 settings, or 85.42%, after Benjamini-Hochberg correction at a 0.05 false-discovery rate.
The central idea is to keep the main recommender frozen while using a request to select how parts of its parameter matrices are masked. That gives the system a way to respond to an immediate instruction without retraining the backbone, which is the study's stated target.
CRAMER mean-pools the request embeddings, uses Gumbel-Top-k gating, decomposes the gates into row and column components, and applies entrywise masks to selected matrices in the frozen Transformer.
The request-to-mask module is trained with a predictive loss plus a KL regularizer. That regularizer encourages the gate distribution to follow a sparsity prior. The paper also says the practical objective is not a strict ELBO. It is a variationally inspired surrogate aligned with hard-sparsity masking.
The test behind the result
The evaluation used four processed datasets: ReDial, with 36,460 interactions; KuaiSAR, with 260,243; Beauty, with 122,485; and CDs&Vinyl, with 141,213. KuaiSAR, Beauty and CDs&Vinyl were randomly downsampled. The comparisons used SASRec and BERT4Rec as the recommender backbones.
For each metric calculation, one ground-truth positive was paired with 100 randomly sampled items that the user had not interacted with. Reported values were averaged over five runs, and significance marking used Benjamini-Hochberg correction at a false-discovery rate of 0.05 across 48 t-tests.
The reported ranking gains
On ReDial and KuaiSAR, CRAMER had higher N@10 values than REARANK with both backbones. With SASRec, the scores were 0.428 versus 0.414 on ReDial and 0.436 versus 0.409 on KuaiSAR. With BERT4Rec, they were 0.451 versus 0.388 on ReDial and 0.434 versus 0.416 on KuaiSAR.
The same pattern appeared on Beauty and CDs&Vinyl. With SASRec, CRAMER scored 0.489 versus REARANK's 0.474 on Beauty and 0.472 versus 0.454 on CDs&Vinyl. With BERT4Rec, the corresponding scores were 0.396 versus 0.379 and 0.487 versus 0.423.
A lighter layer of adaptation
Efficiency was another reported advantage. Under identical settings, CRAMER added 0.018 seconds of runtime and 1,355.6 MiB of GPU memory per SASRec request, compared with REARANK's 9.256 seconds and 9,824.7 MiB.
Under BERT4Rec, CRAMER added 0.021 seconds and 1,408.1 MiB per request. REARANK added 9.184 seconds and 9,412.5 MiB under the same comparison. Together, the reported figures show a much smaller added cost for CRAMER in these tests.
A separate scaling study tested BERT4Rec on KuaiSAR at 4, 6, 8 and 10 layers. At 4 layers, NDCG@10 was 0.434 with CRAMER versus 0.366 for the vanilla model; at 10 layers, it was 0.462 versus 0.385. Across the tested depths, the runtime ratio fell from 55.3% to 41.5%, while the memory ratio fell from 66.4% to 43.7%.
The boundaries of the result
The paper also examined whether the recommendation mix moved in the direction of a request. Among 100 ReDial users, romance-related movies made up 0.286 of the top-10 list with no request, 0.432 with a clear positive request and 0.135 with a clear opposite request. The analysis was an exploratory interpretability study on a sampled user group.
Those results come with a defined boundary. The study used downsampled versions of three datasets, measured ranking against 100 sampled negatives per positive, and summarized the main comparison with five-run averages and BH-adjusted tests. Its training objective is also a variationally inspired surrogate rather than a strict ELBO. The results should therefore be read as benchmark gains under the tested setup.
The supplied record is labeled a preprint. Its front matter presents arXiv:2608.25370v2, dated 13 September 2026, and identifies the Proceedings of the 43rd International Conference on Machine Learning, PMLR 306, 2026.
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