Dynamic Tokenization via Reinforcement Patching: End-to-end Training and Zero-shot Transfer

Yulun Wu; Sravan Kumar Ankireddy; Samuel Sharpe; Nikita Seleznev; Dehao Yuan; Hyeji Kim; Nam H. Nguyen

cs.LGMar 27, 2026

Dynamic Tokenization via Reinforcement Patching: End-to-end Training and Zero-shot Transfer

Authors: Yulun Wu, Sravan Kumar Ankireddy, Samuel Sharpe, Nikita Seleznev, Dehao Yuan, Hyeji Kim, Nam H. Nguyen

Organizations: Capital One · University of Texas, Austin

Abstract

Efficiently aggregating spatial or temporal horizons to acquire compact representations has become a unifying principle in modern deep learning models, yet learning data-adaptive representations for long-horizon sequence data, especially continuous sequences like time series, remains an open challenge. While fixed-size patching has improved scalability and performance, discovering variable-sized, data-driven patches end-to-end often forces models to rely on soft discretization, specific backbones, or heuristic rules. In this work, we propose Reinforcement Patching (ReinPatch), the first framework to jointly optimize a sequence patching policy and its downstream sequence backbone model using reinforcement learning. By formulating patch boundary placement as a discrete decision process optimized via Group Relative Policy Gradient (GRPG), ReinPatch bypasses the need for continuous relaxations and performs dynamic patching policy optimization in a natural manner. Moreover, our method allows strict enforcement of a desired compression rate, freeing the downstream backbone to scale efficiently, and naturally supports multi-level hierarchical modeling. We evaluate ReinPatch on time-series forecasting datasets, where it demonstrates compelling performance compared to state-of-the-art data-driven patching strategies. Furthermore, our detached design allows the patching module to be extracted as a standalone foundation patcher, providing the community with visual and empirical insights into the segmentation behaviors preferred by a purely performance-driven neural patching strategy.

Figures & tables

Method	Data- driven	Variable size	Model- based	End-to- end	Backbone- agnostic	Continuous. exploration	Inductive bias	Data type- specific	Hierarchical modeling
Static Patching [ 14 ]	✗	✗	✗	✗	✓	n/a	✗	✗	✓
Entropy Patching [ 15 ]	✓	✓	✓	✗	✓	n/a	✓	✓	✗
TimeSqueeze’s Compression [ 2 ]	✓	✓	✗	✗	✓	n/a	✓	✓	✗
H-Net’s Chunking [ 8 ]	✓	✓	✓	✓	✗	✓	✓	✗	✓
Periodic Patching [ 21 ]	✓	✗	✗	✗	✓	n/a	✓	✓	✗
Selective Patching [ 24 ]	✓	✗	✓	✓/✗	✗	✓	✓	✗	✗

Table 1: Qualitative comparison of ReinPatch with existing patching strategies across key design properties. ✓ : satisfied, ✗ : not satisfied, ✓/✗ : partially satisfied, n/a: not applicable.

Datasets	ETTh1		ETTh2		ETTm1		ETTm2		Weather		Electricity
Metrics	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
Static Patching [ 14 ]	0.415	0.434	0.337	0.385	0.357	0.388	0.272	0.326	0.230	0.268	0.192	0.299
Selective Patching [ 24 ]	0.416	0.433	0.331	0.384	0.361	0.391	0.269	0.326	0.232	0.271	0.191	0.298
Periodic Patching [ 21 ]	0.408	0.431	0.338	0.387	0.362	0.392	0.274	0.327	0.230	0.268	0.187	0.292
Entropy Patching [ 1 ]	0.430	0.438	0.339	0.389	0.359	0.387	0.283	0.334	0.231	0.269	0.170	0.268
TimeSqueeze’s Compression [ 2 ]	0.422	0.438	0.338	0.388	0.362	0.385	0.272	0.329	0.230	0.268	0.176	0.277

Table 2: Multivariate forecasting average results with forecasting horizons

H\in\{96,192,336,720\}

for the datasets. Red : the best, Blue : the 2nd best. Full results are available in Table 5 of Appendix B.3 .

Figure 1: MSE at forecasting horizon $H=96$ across various look-back windows. While entropy patching exhibits a competitive for short contexts, it struggles to scale as the input context expands, whereas ReinPatch maintains strong forecasting performance across most datasets and context lengths.

Panel of Figure 2 — Figure 2: Representative patch segmentations sampled from the foundation patcher trained with a minimum compression rate of 8. The learned boundaries naturally concentrate around major transitions while keeping short neighborhoods of local temporal structure together.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: Sequence backbone model. Note that this illustration is for contextual tasks. For generative (i.e. causal) tasks, patches are right-shifted by one (and a special token embedding is added to the beginning of the sequence) before upsampling.

Dataset	Domain	Frequency	Lengths	Dim	Split	Description
ETTh1	Electricity	1 hour	14,400	7	6:2:2	Power transformer 1, comprising seven indicators such as oil temperature and useful load
ETTh2	Electricity	1 hour	14,400	7	6:2:2	Power transformer 2, comprising seven indicators such as oil temperature and useful load
ETTm1	Electricity	15 mins	57,600	7	6:2:2	Power transformer 1, comprising seven indicators such as oil temperature and useful load
ETTm2	Electricity	15 mins	57,600	7	6:2:2	Power transformer 2, comprising seven indicators such as oil temperature and useful load
Weather	Environment	10 mins	52,696	21	7:1:2	Recorded every day for the whole year 2020, which contains 21 meteorological indicators
Electricity	Electricity	1 hour	26,304	321	7:1:2	Electricity records the electricity consumption in kWh every 1 hour from 2012 to 2014

Appendix

Table 3: Statistics of datasets.

Method	Encoder ( $\times$ levels)	Decoder ( $\times$ levels)	Forecasting Backbone
Static Patching	Transformer $\times 1$	Transformer $\times 1$	Transformer
Selective Patching	Transformer $\times 1$	Transformer $\times 1$	Transformer
Periodic Patching	Transformer $\times 1$	Transformer $\times 1$	Transformer
Entropy Patching	Transformer $\times 1$	Transformer $\times 1$	Transformer
TimeSqueeze’s Compression	SSM $\times 1$	SSM $\times 1$	Transformer
H-Net’s Chunking	SSM $\times 2$	SSM $\times 2$	Transformer

Appendix

Table 4: Architecture components of different forecasting methods.

Models		ReinPatch (foundation) [Ours]		ReinPatch (end-to-end) [Ours]		Static Patching [ 14 ]		Selective Patching [ 24 ]		Periodic Patching [ 21 ]		Entropy Patching [ 1 ]		TimeSqueeze’s Compression [ 2 ]		H-Net’s Chunking [ 8 ]
Metrics	H	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
ETTh1	96	0.360 $\pm$ 0.008	0.400 $\pm$ 0.006	0.374 $\pm$ 0.007	0.405 $\pm$ 0.006	0.367	0.404	0.368	0.403	0.365	0.403	0.382	0.405	0.378	0.405	0.391	0.409
	192	0.394 $\pm$ 0.003	0.419 $\pm$ 0.004	0.410 $\pm$ 0.016	0.423 $\pm$ 0.009	0.416	0.428	0.416	0.428	0.421	0.432	0.435	0.434	0.423	0.429	0.436	0.437
	336	0.395 $\pm$ 0.007	0.423 $\pm$ 0.005	0.397 $\pm$ 0.007	0.424 $\pm$ 0.006	0.416	0.431	0.419	0.434	0.399	0.424	0.430	0.443	0.417	0.438	0.439	0.450
	720	0.440 $\pm$ 0.019	0.455 $\pm$ 0.010	0.441 $\pm$ 0.016	0.461 $\pm$ 0.011	0.462	0.473	0.460	0.467	0.449	0.465	0.471	0.470	0.467	0.478	0.485	0.483
	Avg.	0.397 $\pm$ 0.008	0.424 $\pm$ 0.005	0.405 $\pm$ 0.008	0.428 $\pm$ 0.004	0.415	0.434	0.416	0.433	0.408	0.431	0.430	0.438	0.422	0.438	0.438	0.445

Appendix

Table 5: Multivariate forecasting full results with forecasting horizons

H\in\{96,192,336,720\}

for the datasets. Red : the best, Blue : the 2nd best. The standard deviation of SRSNet calculated through 3 random seeds can also be reported.

Explore similar work

TimeSqueeze: Dynamic Patching for Efficient Time Series Forecasting
Mar 11, 2026Sravan Kumar Ankireddy, Nikita Seleznev, Nam H. Nguyen +4Time Series Forecasting Time Series Transformers
University of Texas at Austin, Austin, TX, USA · Capital One, USA · University of Maryland, College Park, MD, USA
Adaptive Patching Is Harder Than It Looks For Time-Series Forecasting
Jun 2, 2026Federico Zucchi, Yi Xie, Chao Zhang +2Long-Term Time Series Forecasting Time Series Forecasting
ICube, University of Strasbourg, Illkirch-Graffenstaden, France · Cephalgo, Strasbourg, France · Technical University of Munich +2
IPatch: A Multi-Resolution Transformer Architecture for Robust Time-Series Forecasting
Mar 25, 2026Aymane Harkati, Moncef Garouani, Olivier Teste +2Time Series Forecasting Multivariate Time Series Forecasting
CCPS Laboratory, ENSAM, Hassan II University, Casablanca, 20670, Morocco · IRIT, UMR 5505 CNRS, Université Toulouse Capitole, Toulouse, 31000, France · IRIT, UMR 5505 CNRS, Université Toulouse, UT2J, Toulouse, 31000, France

Mar 11, 2026J/K move · Enter open · S save

TimeSqueeze: Dynamic Patching for Efficient Time Series Forecasting

Sravan Kumar Ankireddy, Nikita Seleznev, Nam H. Nguyen, Yulun Wu+3

University of Texas at Austin, Austin, TX, USA · Capital One, USA · University of Maryland, College Park, MD, USA

Time Series Forecasting Time Series Transformers Time-Series Foundation Models

Open paper AskPDF arXiv

Transformer-based time series foundation models face a fundamental trade-off in choice of tokenization: point-wise embeddings preserve temporal fidelity but scale poorly with sequence length, whereas fixed-length patching improves efficiency by imposing uniform boundaries that may disrupt natural transitions and blur informative local dynamics. In order to address these limitations, we introduce TimeSqueeze, a dynamic patching mechanism that adaptively selects patch boundaries within each sequence based on local signal complexity. TimeSqueeze first applies a lightweight state-space encoder to extract full-resolution point-wise features, then performs content-aware segmentation by allocating short patches to information-dense regions and long patches to smooth or redundant segments. This variable-resolution compression preserves critical temporal structure while substantially reducing the token sequence presented to the Transformer backbone. Specifically for large-scale pretraining, TimeSqueeze attains up to 20x faster convergence and 8x higher data efficiency compared to equivalent point-token baselines. Experiments across long-horizon forecasting benchmarks show that TimeSqueeze consistently outperforms comparable architectures that use either point-wise tokenization or fixed-size patching.

Abstract

Explore similar work

TimeSqueeze: Dynamic Patching for Efficient Time Series Forecasting

Adaptive Patching Is Harder Than It Looks For Time-Series Forecasting

IPatch: A Multi-Resolution Transformer Architecture for Robust Time-Series Forecasting