CRAMER: Control via Request-Aware Masking for Editing Recommenders

Zhiyuan Julian Su; Naihe Feng; Zhen Luther Qin; Ga Wu

cs.IRAug 26, 2026

CRAMER: Control via Request-Aware Masking for Editing Recommenders

Authors: Zhiyuan Julian Su, Naihe Feng, Zhen Luther Qin, Ga Wu

Organizations: Gaoling School of Artificial Intelligence, Renmin University of China · Faculty of Computer Science, Dalhousie University

Abstract

Sequential recommendation models, while powerful, have limited flexibility in responding to immediate user requests, making it difficult to adapt their recommendations to the user's timely interests. Unfortunately, existing user request adaptation methods often incur high computational overhead due to either 1) retraining the entire backbone network or 2) leveraging the inference ability of large language models (a.k.a. prompt engineering), limiting their applicability in large-scale recommendation services. This paper presents Control via Request-Aware Masking for Editing Recommenders (CRAMER), a framework that takes users' natural-language requests to immediately change sequential recommendation models' behavior. Specifically, inspired by the model control theory, CRAMER treats user requests as control signals to modulate frozen backbone parameters through masking, achieving instant adaptation to diverse requests while avoiding costly retraining. Experiments on multiple large-scale benchmark datasets show that CRAMER outperforms four state-of-the-art request-aware baselines across multiple recommendation metrics while achieving minimal overhead. Moreover, the proposed framework exhibits enhanced controllability and cross-domain adaptability, establishing a new paradigm for request-aware sequential recommendation.

Figures & tables

Figure 1: The overview of the proposed CRAMER framework. The figure shows the whole process of CRAMER converting a natural-language request into masks and controlling the sequential recommender (frozen sequential recommender backbone). The gray area with a pink dashed border represents the “Row–Column Gating Masks” paragraph in Section 3.3 .

Method	ReDial						KuaiSAR
Method	H@10	H@20	N@10	N@20	M@10	M@20	H@10	H@20	N@10	N@20	M@10	M@20
SASRec	0.426	0.573	0.373	0.410	0.344	0.354	0.430	0.601	0.346	0.389	0.306	0.318
+Query-SeqRec	0.450	0.596	0.391	0.429	0.350	0.361	0.451	0.567	0.348	0.378	0.313	0.322
+BLaIR	0.447	0.582	0.392	0.426	0.287	0.296	0.479	0.612	0.408	0.443	0.293	0.303
+LLM-ESR	0.516	0.666	0.385	0.423	0.323	0.334	0.496	0.628	0.392	0.427	0.343	0.351
+REARANK	0.549	0.684	0.414	0.449	0.408	0.417	0.538	0.645	0.409	0.437	0.366	0.374

Table 1: Overall results of four baselines and our CRAMER. H@

k

, N@

k

, and M@

k

denote HR@

k

, NDCG@

k

, and MRR@

k

, respectively (averaged over five runs). For each setting, the boldface refers to the highest result, and the underline indicates the second best result. “*” marks statistically significant improvements after BH procedure (FDR = 0.05) across all 48 t-tests.

Figure 2: Sensitivity of CRAMER to drop ratio $\rho$ , evaluated using NDCG@10. For each setting, five evaluations were performed, the column height shows its average result.

Figure 3: Impact of different PLMs on request encoding, evaluated using NDCG@10. For each setting, five evaluations were performed, the column height shows its average result.

Method	Runtime (s)	GPU Memory (MiB)
SASRec	0.033	2024.1
+Query-SeqRec	+0.021	+1587.5
+BLaIR	+0.029	+1125.4
+LLM-ESR	+0.016	+1236.2
+REARANK	+9.256	+9824.7
+CRAMER (Ours)	+0.018	+1355.6

Table 2: Inference efficiency comparison for SASRec backbone. Runtime and GPU memory usage are measured as average per-request cost under identical settings. “Vanilla Backbone” reports the backbone-only cost, and the remaining rows show the additional overhead introduced by each request-aware method.

Type	Request Text	Avg	Var
$\backslash$	No request	0.286	0.0218
(1)	“I’d like a romantic comedy.”	0.432 $\uparrow$	0.0244
(2)	“Something sweet and heartwarming.”	0.345 $\uparrow$	0.0253
(3)	“I want an offbeat, slow-burn emotional drama.”	0.312 $\uparrow$	0.0371
(4)	“Please avoid romantic movies.”	0.135 $\downarrow$	0.0187
(5)	“Maybe something less focused on love.”	0.204 $\downarrow$	0.0198

Table 3: The proportion of romance-related movies in the top-10 recommendations under different request types (100 users). The first row reports the baseline results without any request.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	#Items	#Inters	Average Chars
ReDial	5207	36460	112.84
KuaiSAR	174895	260243	124.63
Beauty	44977	122485	226.85
CDs&Vinyl	76368	141213	973.21

Appendix

Table 4: Statistics of the processed datasets. “#Items” represents the total number of items, “#Inters” represents the total number of interactions (each one contains a request), and “Average Chars” represents the average number of characters of requests.

Backbone	Dataset	$\rho$	PLM	$\theta_{M}$	$\phi_{t}$	$\lambda_{\text{KL}}$	$\tau$ (0.7 $\to$ 0.3)
SASRec	ReDial	0.20	RoBERTa	$W_{O}$	last	0.4	cosine
	KuaiSAR	0.05	ModernBERT	FFNs+ $W_{O}$	all	0.1	cosine
	Beauty	0.10	RoBERTa	FFNs+ $W_{O}$	last	0.2	cosine
	CDs&Vinyl	0.10	ModernBERT	FFNs+ $W_{O}$	all	0.1	cosine
BERT4Rec	ReDial	0.15	RoBERTa	FFNs	last	0.3	cosine
	KuaiSAR	0.05	ModernBERT	FFNs+ $W_{O}$	all	0.1	cosine

Appendix

Table 5: Optimal hyperparameter settings for each backbone

\times

dataset configuration.

Figure 4: Impact of different scopes of $\theta_{M}$ , evaluated using NDCG@10. For each setting, five evaluations were performed, the column height shows its average result.

Figure 5: Impact of different regimes of $\phi_{t}$ , evaluated using NDCG@10. For each setting, five evaluations were performed, the column height shows its average result.

Figure 6: Sensitivity of CRAMER to weight $\lambda_{\text{KL}}$ , evaluated using NDCG@10. For each setting, five evaluations were performed, the column height shows its average result.

Figure 7: The effect of whether gates are sampled once and shared across the entire batch (1) or sampled independently per instance (0) in the training phase, evaluated using NDCG@10. For each setting, five evaluations were performed, the column height shows its average result.

Figure 8: The effect of different temperature annealing schedule, evaluated using NDCG@10. For each setting, five evaluations were performed, the column height shows its average result.

Method	Runtime (s)	GPU Memory (MiB)
BERT4Rec	0.038	2119.6
+Query-SeqRec	+0.023	+1620.4
+BLaIR	+0.027	+1102.6
+LLM-ESR	+0.018	+1375.3
+REARANK	+9.184	+9412.5
+CRAMER (Ours)	+0.021	+1408.1

Appendix

Table 6: Inference efficiency comparison for BERT4Rec backbone. Runtime and GPU memory usage are measured as average per-request cost under identical settings. “Vanilla Backbone” reports the backbone-only cost, and the remaining rows show the additional overhead introduced by each request-aware method.

BERT Version	Tiny	Mini	Medium	Base
ReDial	0.403	0.411	0.419	0.428
KuaiSAR	0.420	0.425	0.432	0.436
Beauty	0.465	0.474	0.481	0.489
CDs&Vinyl	0.449	0.454	0.460	0.472

Appendix

Table 8: Average NDCG@10 performance (five evaluations) of CRAMER under SASRec when initialized with four BERT versions of different capacity. Across all datasets, performance exhibits a strictly monotonic improvement that aligns with PLM capacity, illustrating that CRAMER faithfully leverages the semantic quality of the request encoder.

Index	User ID	#Inters	Last Item ID
#1	AE25K5V5RESPJ4WKCALB3ZVYYQPQ	11	B000008KJ8
#2	AFE66HHU55NJMALT34HEODVGEPQA	6	B00KNTDO3I
#3	AG2CJZJORAG7SG32SYNTNHICMGOQ	8	B07RF4JVGJ
#4	AGUPFBZ756HTU4YIF4QKQEX3NS2Q	13	B00SFXFCWA
#5	AHQF2VXWQPUBKYR3RMJ6VDFDYUSQ	9	B08L47S144

Appendix

Table 9: Detailed interaction information for the five users selected from the CDs&Vinyl dataset. For each user with a “User ID”, “#Inters” represents the total number of this user’s interactions, and “Last Item ID” represents the ID of the item in this user’s last interaction. Because we use leave-one-out (LOO) evaluation, the item in the last interaction is the ground-truth item in evaluation.

SASRec
Method	#1	#2	#3	#4	#5
$\backslash$	18.2	28.4	27.0	19.0	17.8
Query-SeqRec	16.4	23.2	18.6	23.2	15.0
BLaIR	12.2	20.4	10.4	13.4	13.2
LLM-ESR	14.0	17.0	15.8	15.2	9.4
REARANK	13.6	23.2	11.2	10.0	11.4

Appendix

Table 10: Average ranking of true positive items of users selected from CDs&Vinyl (smaller is better). For each setting, five evaluations were performed, and boldface indicates the best, i.e., lowest, average rank under the same backbone.

Recommender Systems

Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.

See all 502 papers

RAIE: Region-Aware Incremental Preference Editing with LoRA for LLM-based Recommendation
Feb 28, 2026Jin Zeng, Yupeng Qi, Hui Li +4Recommender Systems Large Language Model-Based Recommendation
School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen, Guangdong, China · Department of Computer Science and Technology, Xiamen University, Xiamen, Fujian, China · Artificial Intelligence Research Institute, Shenzhen MSU-BIT University, Shenzhen, Guangdong, China +2
Give Users the Wheel: Towards Promptable Recommendation Paradigm
Feb 21, 2026Fuyuan Lyu, Chenglin Luo, Qiyuan Zhang +6Sequential Recommendation User Preference Modeling
McGill & Mila - Quebec AI Institute Montreal, Canada · Shenzhen Technological University Shenzhen, China · City Univerisity of Hong Kong Hong Kong, China +3
RecRec: Recursive Refinement for Sequential Recommendation
Jul 12, 2026Pervez Shaik, Prosenjit Biswas, Abhinav Thorat +2Sequential Recommendation Iterative Refinement
Sony Research India India

Feb 28, 2026J/K move · Enter open · S save

RAIE: Region-Aware Incremental Preference Editing with LoRA for LLM-based Recommendation

Jin Zeng, Yupeng Qi, Hui Li, Chengming Li+3

School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen, Guangdong, China · Department of Computer Science and Technology, Xiamen University, Xiamen, Fujian, China · Artificial Intelligence Research Institute, Shenzhen MSU-BIT University, Shenzhen, Guangdong, China+2

Recommender Systems Large Language Model-Based Recommendation Incremental Learning User Preference Modeling

Open paper AskPDF arXiv

Large language models (LLMs) are increasingly adopted as the backbone of recommender systems. However, user-item interactions in real-world scenarios are non-stationary, making preference drift over time inevitable. Existing model update strategies mainly rely on global fine-tuning or pointwise editing, but they face two fundamental challenges: (i) imbalanced update granularity, where global updates perturb behaviors unrelated to the target while pointwise edits fail to capture broader preference shifts; (ii) unstable incremental updates, where repeated edits interfere with prior adaptations, leading to catastrophic forgetting and inconsistent recommendations. To address these issues, we propose Region-Aware Incremental Editing (RAIE), a plug-in framework that freezes the backbone model and performs region-level updates. RAIE first constructs semantically coherent preference regions via spherical k-means in the representation space. It then assigns incoming sequences to regions via confidence-aware gating and performs three localized edit operations - Update, Expand, and Add - to dynamically revise the affected region. Each region is equipped with a dedicated Low-Rank Adaptation (LoRA) module, which is trained only on the region's updated data. During inference, RAIE routes each user sequence to its corresponding region and activates the region-specific adapter for prediction. Experiments on two benchmark datasets under a time-sliced protocol that segments data into Set-up (S), Finetune (F), and Test (T) show that RAIE significantly outperforms state-of-the-art baselines while effectively mitigating forgetting. These results demonstrate that region-aware editing offers an accurate and scalable mechanism for continual adaptation in dynamic recommendation scenarios. Our code is available at https://github.com/fengaogao/RAIE.

Abstract

Recommender Systems

Explore similar work

RAIE: Region-Aware Incremental Preference Editing with LoRA for LLM-based Recommendation

Give Users the Wheel: Towards Promptable Recommendation Paradigm

RecRec: Recursive Refinement for Sequential Recommendation