DP-Rec: Towards Dynamic Patching for Efficient Long-Sequence Recommendation

Dwipam Katariya; Thomas Caputo; Akshat Shreemali; Juan Manuel Origgi; Nikita Seleznev; Pranab Mohanty; Kalanand Mishra; Nam Nguyen

cs.LGSep 26, 2026

DP-Rec: Towards Dynamic Patching for Efficient Long-Sequence Recommendation

Authors: Dwipam Katariya, Thomas Caputo, Akshat Shreemali, Juan Manuel Origgi, Nikita Seleznev, Pranab Mohanty, Kalanand Mishra, Nam Nguyen,

Organizations: Capital One, AI Foundations McLean, VA, USA

Abstract

Transformers have redefined sequential recommendation by effectively modeling dynamic user behaviors and long-range dependencies. However, they remain inherently inefficient: standard architectures operate at a fixed rate, allocating comparable computation to every item in a user's history regardless of its information content. This leads to prohibitive computational overhead on long sequences and increased sensitivity to behavioral noise. To address this, practitioners often resort to lossy sequence compression, staged modeling, or truncation. This limits the model's ability to leverage the full context of long histories during inference. Inspired by the recent success of Byte Latent Transformer, we propose DP-Rec, a dynamic latent patching architecture for recommendation. DP-Rec shifts from item-level modeling to patch-level modeling by segmenting interaction sequences using contrastive entropy surprise to identify informative behavioral boundaries. A lightweight patch encoder compresses these temporally contextualized segments into a reduced set of dynamic latent behavior vectors, which are then processed by a larger latent transformer and decoded for next-item prediction. Extensive experiments show that, under constrained computational budgets, DP-Rec scales effectively to long sequences and achieves a superior efficiency-accuracy trade-off over both non-compressed and fixed-size compression baselines.

Figures & tables

Figure 1. Dynamic vs Static Patching. Unlike static patching (green square), DP-Rec adaptively partitions interaction streams at contrastive entropy peaks (e.g., Interstellar, The Conjuring) (red circle). These peaks signify shifts in latent user intent. Optimized threshold $\tau$ (red dashed line) results in variable-length patching, focusing computational resources on informative behavioral boundaries rather than redundant items to maximize recommendation efficiency. A line graph of entropy over a user interaction stream. A dashed line marks threshold tau. Red circles mark high-entropy peaks for Dynamic Patching; a green square marks a Fixed-size Patching interval

Symbol	Meaning
$\mathcal{U}$	Set of users
$\mathcal{V}$	Set of items
$B$	Mini-batch size
$N$	Raw sequence length ( $N_{\mathrm{max}},N_{\mathrm{raw}}$ )
$\mathbf{s}_{u}$	Interaction sequence of user $u$ , $(v_{1},\dots,v_{N})$
$v_{t}$	Item ID at position $t$

Table 1. Notations

Figure 2. Overall architecture of DP-Rec. A frozen detector computes contrastive-entropy scores, enabling an adaptive threshold optimizer to partition sequences into $M$ variable-length behavioral patches. These are contextualized by a temporal local encoder and modeled via a temporal latent transformer before a local decoder projects representations back to the item space for next-item prediction. Fully described in the caption and main text.

Figure 3. Impact of batch size on patch budget adherence (ML-1M, $\hat{M}=50$ ) for $N\geq 100$ . While both configurations maintain the target mean, the Batch Size $BS=128$ distribution shows the effect of cross-sequence budget reallocation within mini-batches, contrasted with the high-precision adherence of the $BS=1$ case Histogram of patches per sequence. Blue bars (batch size 1) show low variance; yellow bars (batch size 128) show higher variance. A dashed vertical line marks the target patch count controlled by tau.

Dataset	$\mathcal{U}$	$\mathcal{V}$	# Actions	Avg. Actions / User
KuaiRand ( Gao et al., 2022 )	27,285	13,222,181	122,052,626	4,473
ML-1M ( Harper and Konstan, 2015 )	6,040	3,416	999,611	165
ML-10M-L ( Harper and Konstan, 2015 )	3,683	10,632	3,114,710	844

Table 2. Dataset Summary

		Fixed-Horizon Baselines			Compression-Based Models
Dataset	Metric	GRU4Rec ( Hidasi et al., 2015a )	SASRec ( Kang and McAuley, 2018 )	HSTU ( Zhai et al., 2024 )	LONGER ( Chai et al., 2025 )	DP-Rec	Rel Impr.
ML-1M ( $N_{\text{max}}=50,N_{\text{raw}}=200$ )
	NDCG@10	$0.1961\pm 0.0023$	$0.2176\pm 0.0031$	$0.2078\pm 0.0051$	$0.2022\pm 0.0042$	$\mathbf{0.2523\pm 0.0017}$	+16.0%
	NDCG@20	$0.2308\pm 0.0026$	$0.2525\pm 0.0029$	$0.2410\pm 0.0049$	$0.2355\pm 0.0035$	$\mathbf{0.2883\pm 0.0013}$	+14.2%
	HR@5	$0.2329\pm 0.0057$	$0.2628\pm 0.0045$	$0.2535\pm 0.0068$	$0.2445\pm 0.0054$	$\mathbf{0.3092\pm 0.0024}$	+17.7%
	HR@10	$0.3545\pm 0.0025$	$0.3846\pm 0.0055$	$0.3696\pm 0.0080$	$0.3637\pm 0.0039$	$\mathbf{0.4430\pm 0.0044}$	+15.2%

Table 3. Test results under a fixed FLOPs budget matched to best-validation SASRec, reported as mean

\pm

95% CI across 10 seeds. Bold and underline denote the best overall and baseline performers, and Rel Impr. measures relative gain over the top baseline.

Panel of Figure 4 — Figure 4. Efficiency trends for budgeted inference FLOPs. We plot NDCG@10 against total inference FLOPs ( $\log_{10}$ scale). DP-Rec (red line) consistently defines the upper-left frontier, demonstrating its ability to either (i) achieve higher recommendation quality at the same inference cost, or (ii) reach equivalent quality levels with a significantly reduced computational footprint compared to baselines. Line plot showing the Pareto curve for each model. The y-axis is Normalized Discounted Cumulative Gain at 10 (NDCG@10), the x-axis is total inference Floating Point Operations on a log base 10 scale. Each model is a different color with a red line for DP-Rec above the others, indicating better accuracy when controlling for compute

$d_{\mathrm{loc}}$	$L_{\mathrm{loc}}$	#Params	ML-1M		ML-10M-L		KuaiRand
		(non-emb)	NDCG@10 $\uparrow$	TFLOPs $\downarrow$	NDCG@10 $\uparrow$	TFLOPs $\downarrow$	NDCG@10 $\uparrow$	TFLOPs $\downarrow$
32	1	137,920	0.2238	0.0284 ( $1.00\times$ )	0.2051	0.0740 ( $1.00\times$ )	0.4368	0.5431 ( $1.00\times$ )
16	1	99,616	0.1909 (-14.70%)	0.0100 ( $0.35\times$ )	0.1927 (-6.05%)	0.0235 ( $0.32\times$ )	0.4118 (-5.72%)	0.1729 ( $0.32\times$ )
64	1	224,768	0.2389 (+6.75%)	0.0989 ( $3.48\times$ )	0.2108 (+2.78%)	0.2663 ( $3.60\times$ )	0.4405 (+0.85%)	1.9523 ( $3.59\times$ )
32	2	159,168	0.2219 (-0.85%)	0.0543 ( $1.91\times$ )	0.2095 (+2.15%)	0.1447 ( $1.96\times$ )	0.4383 (+0.34%)	1.0606 ( $1.95\times$ )
32	3	180,416	0.2145 (-4.16%)	0.0800 ( $2.82\times$ )	0.2010 (-2.00%)	0.2153 ( $2.91\times$ )	0.4407 (+0.89%)	1.5782 ( $2.91\times$ )

Table 4. Ablation on local module capacity. Bold denotes the base configuration (

1.00\times

compute). Parentheses denote relative change (

\Delta\%

) for NDCG and compute scale for TFLOPs.

Variant	ML-1M	ML-10M-L	KuaiRand
DP-Rec (baseline)	0.2238	0.2051	0.4368
w/o Time-RoPE	0.2018 (-9.83%)	0.1933 (-5.75%)	0.4281 (-1.99%)
w/o cross-attn	0.2424 (+8.31%)	0.2026 (-1.22%)	0.4444 (+1.74%)
$k=2$	0.2158 (-3.57%)	0.2016 (-1.71%)	0.4389 (+0.48%)
$k=32$	0.2302 (+2.86%)	0.2012 (-1.90%)	0.4344 (-0.55%)
$w=8$	0.1956 (-12.60%)	0.1916 (-6.58%)	0.4006 (-8.29%)

Table 5. Architectural ablations under extreme compression (

\hat{M}=2

). Parentheses denote relative change (

\Delta\%

) from the base configuration.

Dataset	Random	Fixed	CE
ML-1M	0.2400	0.2437	0.2540
ML-10M-L	0.1946	0.1982	0.2029
KuaiRand	0.4588	0.4634	0.4761

Table 6. Patching-scheme ablation (Random, Fixed, Contrastive-Entropy) at

\hat{M}=50

for ML-1M and

\hat{M}=100

for ML-10M-L and KuaiRand , varying only the boundary selection rule. Single-seed NDCG@10 scores are reported; bold marks the best per dataset.

Target Patches ( $\hat{M}$ )	ML-1M	ML-10M-L	KuaiRand
2	0.2346	0.2050	0.4405
5	0.2446	0.2026	0.4414
10	0.2414	0.2070	0.4504
50	0.2499	0.1996	0.4578
100	0.2486	0.1949	0.4502
250	–	0.2088	0.4417

Table 7. Mean test NDCG@10 vs. target patch budget

\hat{M}

, at fixed

N_{\text{raw}}=50

for ML-1M and

N_{\text{raw}}=100

for ML-10M-L and KuaiRand . Cross-attention is disabled to isolate patch budget sensitivity from cross-attention capacity effects. Bold marks best per dataset.

Figure 5. Qualitative analysis of contrastive entropy surprise (red) and item views (blue) for ML-1M . Entropy remains low during stable genre engagement but spikes sharply at out-of-distribution transitions, demonstrating how DP-Rec adaptively allocates boundaries to salient behavioral shifts. Table of viewed movies with title, genre, entropy, and views columns shown as inline bars. Green brackets group genre-stable rows with low entropy; red brackets mark genre shifts with high entropy.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 6 — Figure 6. Recommendation Quality (NDCG@10) by Item Popularity across three benchmark datasets.

Panel of Figure 7 — Figure 7. Recommendation Quality(NDCG@10) by Time Lag between the last item and the item in the test to be ranked.

Configuration	Fixed-Horizon Models			Compression-Based Models
	SASRec ( Huang, 2020 )	HSTU ( MobileTeleSystems, 2024 )	GRU4Rec ( MobileTeleSystems, 2024 )	LONGER	DP-Rec
Blocks ( $L$ )	3	3	$\{1,3\}$	$L_{\mathrm{inner}},L_{\mathrm{outer}}\in\{1,3\}$	$L_{\mathrm{enc}}=1,L_{\mathrm{dec}}=1,L_{\mathrm{lat}}\in\{1,3\}$
Hidden size ( $d$ )	$\{32,64\}$	64	64	$\{16,32,64\}$	$\{16,32,64\}$
Seq. length ( $N$ )	$\{25,\dots,500\}$	Matched	$\{50,100\}$	$\{100,\dots,500\}$	$\{100,\dots,500\}$
Patch Budget ( $\hat{M}$ )	–	–	–	$\{2,\dots,250\}$	$\{2,\dots,250\}$
Pool Initialization	True	True	True	True	True

Appendix

Table 8. Summary of experimental configurations and hyperparameter search spaces. For direct efficiency comparisons, Iso-capacity variants utilize a

1/1/1

block structure (

d=64

) to parameter-match SASRec (

d=64,L=3

Model	FLOPs	vs. SASRec	Complexity
SASRec	$74{,}419{,}200$	$1.00\times$	$O(n^{2}h\ell)$
HSTU	$81{,}861{,}120$	$1.10\times$	$O(n^{2}h\ell)$
LONGER	$26{,}310{,}400$	$0.35\times$	$O(nkd_{\mathrm{loc}}L_{\mathrm{loc}}+G^{2}dL_{\mathrm{lat}})$
DP-Rec	$42{,}421{,}120$	$0.57\times$	$O(nw\,d_{\mathrm{loc}}\cdot 2L_{\mathrm{loc}}+M^{2}d\,L_{\mathrm{lat}})$

Appendix

Table 9. Concrete FLOPs at

n=200

d_{\mathrm{ff}}=4

. SASRec/HSTU:

d=64

\ell=3

. LONGER:

d=d_{\mathrm{loc}}=64

L_{\mathrm{lat}}=L_{\mathrm{loc}}=1

k=4

G=50

. DP-Rec:

d=d_{\mathrm{loc}}=64

L_{\mathrm{lat}}=L_{\mathrm{loc}}=1

w=32

d_{\mathrm{bbd}}=8

L_{\mathrm{bbd}}=1

w_{e}=32

M=10

Sequential Recommendation

Sequential recommendation predicts what a user will choose next from ordered interactions, adapting to changing interests and context.

See all 136 papers

SequenceO1: End-to-End Ultra-Long (100K) Sequence Modeling in Recommendation with Low-Rank Caching
Sep 8, 2026Lin Guan, Jia-Qi Yang, Zhishan Zhao +15Sequential Recommendation Long-Context Modeling
ByteDance
HyTRec: A Hybrid Temporal-Aware Attention Architecture for Long Behavior Sequential Recommendation
Feb 20, 2026Lei Xin, Yuhao Zheng, Ke Cheng +3Sequential Recommendation Temporal Attention
Shanghai Dewu Information Group · Wuhan University · USTC +1
HiSAC: Hierarchical Sparse Activation Compression for Ultra-long Sequence Modeling in Recommenders
Feb 24, 2026Kun Yuan, Junyu Bi, Daixuan Cheng +5Sequential Recommendation Hierarchical Representation Learning
Alibaba Group · Beijing, China · Renmin University

Sep 8, 2026J/K move · Enter open · S save

SequenceO1: End-to-End Ultra-Long (100K) Sequence Modeling in Recommendation with Low-Rank Caching

Lin Guan, Jia-Qi Yang, Zhishan Zhao, Jiaqi Huang+14

ByteDance

Sequential Recommendation Long-Context Modeling User Preference Modeling

Open paper AskPDF arXiv

Modeling long-term user behavior is central to sequential recommendation and billion-scale industrial recommender systems, yet production ranking models operate under strict latency, memory, communication, and training-throughput constraints. At the 100K scale, the challenge extends beyond attention complexity: raw sequence features must be stored, transferred, and repeatedly processed during training and online serving. Existing approaches based on history truncation, multi-stage behavior retrieval, compressed lifelong histories, or train-short/infer-long extrapolation either weaken end-to-end optimization or retain substantial length-dependent cost. We present SequenceO1, an end-to-end framework for ultra-long user behavior sequence modeling, deployed at full traffic on Douyin with histories of up to 100K interactions. SequenceO1 follows a compress-then-reason design. Its Sketch Attention (SA) uses learnable prototypes and prototype-wise normalization to compress the raw history into a fixed-size, target-agnostic user representation. Target-conditioned Stacked Target-to-History Cross Attention (STCA) then models complementary time scales: a recent 10K suffix for short-term interests and the compact sketch for long-term preferences. To make training and inference practical, SequenceO1 combines low-rank user representation caching, multi-request user-level batching, pipeline lift, and a fused FlashSA kernel to amortize feature storage, communication, and computation across targets, training instances, and consecutive requests. Production experiments show consistent offline and online gains, while the compact cached sketch retains most of the benefit of directly scaling end-to-end sequence ranking to 100K. These results provide a practical model-system approach to efficient attention, sequence compression, and scalable long-sequence and long-context recommendation systems.

Abstract

Sequential Recommendation

Explore similar work

SequenceO1: End-to-End Ultra-Long (100K) Sequence Modeling in Recommendation with Low-Rank Caching

HyTRec: A Hybrid Temporal-Aware Attention Architecture for Long Behavior Sequential Recommendation

HiSAC: Hierarchical Sparse Activation Compression for Ultra-long Sequence Modeling in Recommenders