Research map / Time series / Foundation models for time series

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

NeurIPS 2026 · Time series › Foundation models for time series · approach: LLM-based time series forecasting · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ReinPatch, which learns variable-size patch boundaries with a group-relative policy gradient (GRPG) over discrete boundary decisions, jointly with a downstream forecasting backbone, and evaluates end-to-end and zero-shot foundation-patcher variants on long-horizon time-series forecasting.

Problem

Learning data-adaptive, variable-size patch boundaries end-to-end for long sequences (especially continuous time series) without soft discretization, backbone-specific designs, or heuristic/external rules.

Approach

Treat patch boundary placement as a discrete action chosen by a separate lightweight transformer policy, optimized with group-relative policy gradient using the negative forecasting loss as terminal reward (group mean/std as baseline, no critic), with the backbone as the environment; a hard minimum compression rate is enforced by merging trailing patches.

Key results

• Zero-shot foundation patcher achieves best average MSE on 4 of 6 datasets (e.g., ETTh1 0.397 vs 0.408 for periodic patching; Electricity 0.168 vs 0.170 for entropy patching) and reports 17 first-place results among 24 configurations.
• End-to-end ReinPatch reports an average MSE reduction of 4.4% over entropy patching, peaking at 7.7% on ETTh1.
• H-Net chunking performs worst on the two ETTh subsets (ETTh1 avg MSE 0.438).

Surprising findings

• The zero-shot foundation patcher pre-trained on UTSD outperforms end-to-end training of ReinPatch on the target datasets.
• The learned policy places boundaries that wrap major turning points rather than splitting at each one, contrary to common shifting-point heuristics.

Limitations

• Foundation patcher pre-trained only on univariate series; multivariate pre-training not explored.
• Inference-time compression rate control for generative (causal) tasks is not solved; the one-step MDP is not applied to causal settings.
• Evaluation limited to time series forecasting; no language or other domain experiments.

Experimental setup

Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity UTSD

Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity

Metrics: MSE MAE

Baselines: Static Patching (PatchTST) Selective Patching (SRSNet) Periodic Patching (LightGTS) Entropy Patching (EntroPE) TimeSqueeze compression H-Net chunking

Builds on: PatchTST GRPO H-Net Byte Latent Transformer

Closest papers by meaning

• You Can Learn Tokenization End-to-End with Reinforcement Learning (ICML 2026 · similarity 0.86)
• Multi-Patch Prediction: Adapting Language Models for Time Series Representation Learning (ICML 2024 · similarity 0.86)
• Beyond Next Token Prediction: Patch-Level Training for Large Language Models (ICLR 2025 · similarity 0.85)
• Understanding Layer Patching in Model Size Interpolation (NeurIPS 2026 · similarity 0.84)
• DiffStitch: Boosting Offline Reinforcement Learning with Diffusion-based Trajectory Stitching (ICML 2024 · similarity 0.84)
• Reinforced Fast Weights via Next-Sequence Prediction (NeurIPS 2026 · similarity 0.84)