An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling
Pith reviewed 2026-05-11 19:30 UTC · model grok-4.3
The pith
A simple convolutional architecture outperforms LSTMs on diverse sequence tasks while showing longer effective memory.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The authors evaluate generic convolutional and recurrent architectures across a broad range of standard sequence modeling tasks used to benchmark recurrent networks. Their results indicate that a simple convolutional architecture outperforms canonical recurrent networks such as LSTMs, while demonstrating longer effective memory. They conclude that the common association between sequence modeling and recurrent networks should be reconsidered, and that convolutional networks should be regarded as a natural starting point for sequence modeling tasks.
What carries the argument
A simple convolutional architecture that processes sequences in parallel and captures long-range dependencies without recurrence.
Load-bearing premise
The chosen tasks and datasets represent general sequence modeling challenges and both architectures are implemented and compared fairly without hidden advantages.
What would settle it
A standard sequence modeling benchmark where an LSTM records higher accuracy or F1 score than the convolutional model, or where the convolutional model's effective memory span proves shorter than the LSTM's.
read the original abstract
For most deep learning practitioners, sequence modeling is synonymous with recurrent networks. Yet recent results indicate that convolutional architectures can outperform recurrent networks on tasks such as audio synthesis and machine translation. Given a new sequence modeling task or dataset, which architecture should one use? We conduct a systematic evaluation of generic convolutional and recurrent architectures for sequence modeling. The models are evaluated across a broad range of standard tasks that are commonly used to benchmark recurrent networks. Our results indicate that a simple convolutional architecture outperforms canonical recurrent networks such as LSTMs across a diverse range of tasks and datasets, while demonstrating longer effective memory. We conclude that the common association between sequence modeling and recurrent networks should be reconsidered, and convolutional networks should be regarded as a natural starting point for sequence modeling tasks. To assist related work, we have made code available at http://github.com/locuslab/TCN .
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript conducts a systematic empirical evaluation of generic convolutional and recurrent architectures for sequence modeling on standard benchmark tasks. It claims that a simple convolutional architecture outperforms canonical recurrent networks such as LSTMs across a diverse range of tasks and datasets while demonstrating longer effective memory, and concludes that convolutional networks should be regarded as a natural starting point for sequence modeling tasks. Code is made available at a GitHub repository.
Significance. If the reported results hold, the work would be significant in challenging the default association of sequence modeling with recurrent networks and in providing evidence for convolutional architectures as a competitive or superior alternative, supported by the public code release which enables reproducibility and independent verification.
major comments (1)
- [Abstract] Abstract: The central claims of outperformance over LSTMs and longer effective memory are presented as direct results without any quantitative metrics, tables, task descriptions, implementation details, hyperparameter protocols, or statistical tests in the provided manuscript; these elements are load-bearing for assessing whether the data support the conclusion that convolutional networks should replace recurrent ones as the default.
Simulated Author's Rebuttal
We thank the referee for their review and for highlighting the need for clarity on how the abstract relates to the supporting evidence in the manuscript. We address the major comment below.
read point-by-point responses
-
Referee: [Abstract] Abstract: The central claims of outperformance over LSTMs and longer effective memory are presented as direct results without any quantitative metrics, tables, task descriptions, implementation details, hyperparameter protocols, or statistical tests in the provided manuscript; these elements are load-bearing for assessing whether the data support the conclusion that convolutional networks should replace recurrent ones as the default.
Authors: The abstract is a concise summary of the paper's primary conclusions, following standard academic conventions that reserve detailed evidence for the main text. The full manuscript supplies all requested elements: task and dataset descriptions appear in Section 4, the experimental protocol including implementation details and hyperparameter search is given in Section 5, quantitative results with direct LSTM comparisons are reported in Tables 1--3 and Figures 2--4, and the longer effective memory analysis is presented in Section 5.3 with supporting dilation and receptive-field experiments. Performance differences are shown consistently across multiple independent runs and diverse benchmarks, supporting the claim that the convolutional architecture is a competitive starting point. The public code release enables independent verification of all reported numbers. revision: no
Circularity Check
No significant circularity
full rationale
The paper is a purely empirical study that performs a systematic comparison of convolutional and recurrent architectures on standard sequence modeling benchmarks. No derivations, first-principles results, fitted parameters presented as predictions, or self-referential equations appear in the abstract or described methodology. Claims rest on direct experimental measurements against external tasks and datasets, with code released for reproducibility. No load-bearing steps reduce to inputs by construction.
Axiom & Free-Parameter Ledger
Forward citations
Cited by 60 Pith papers
-
BAH Dataset for Ambivalence/Hesitancy Recognition in Videos for Digital Behavioural Change
Introduces the BAH dataset with 1,427 annotated videos for multimodal recognition of ambivalence/hesitancy in digital behavior change contexts.
-
Efficiently Modeling Long Sequences with Structured State Spaces
S4 is an efficient state space sequence model that captures long-range dependencies via structured parameterization of the SSM, achieving state-of-the-art results on the Long Range Arena and other benchmarks while bei...
-
ConTex: Reformulating Counterfactual Generation For Time Series Forecasting
ConTex learns a global intervention strategy via a decomposed temporal-conditional encoder architecture to generate consistent, sparse counterfactuals for time series models in a single forward pass.
-
OpenGlass: Ultra-Low-Power On-Device AI Eyewear with Event-based Vision
OpenGlass is an open-source smart glasses platform using event-based vision and event-driven power management to achieve 11.5 hours of continuous on-device ML on a 200 mAh battery, demonstrated with 83.94% cross-subje...
-
Scale-Equivariant Generative Forecasting: Weight-Tied Dilated Convolutions, Wavelet Scattering Inputs, and Spectral-Consistency Training for Self-Similar Time Series
Presents SE-WaveNet with weight-tied dilated convolutions plus wavelet and spectral components that reproduces empirical scaling collapse on financial returns while using L times fewer convolutional parameters.
-
U-STS-LLM A Unified Spatio-Temporal Steered Large Language Model for Traffic Prediction and Imputation
U-STS-LLM uses a spatio-temporally steered LLM with dynamic attention bias generation to achieve state-of-the-art results on long-horizon traffic forecasting and high-missing-rate imputation while remaining parameter-...
-
Clin-JEPA: A Multi-Phase Co-Training Framework for Joint-Embedding Predictive Pretraining on EHR Patient Trajectories
Clin-JEPA supplies a multi-phase co-training method for JEPA pretraining on EHR trajectories that achieves converging latent rollouts and improved multi-task AUROC on MIMIC-IV data.
-
TailedTS: Benchmark Dataset for Heavy-Tailed Time Series Prediction and Periodicity Quantification
TailedTS supplies 24.69 billion Wikipedia page-view records as a public benchmark for heavy-tailed time series forecasting and periodicity analysis, revealing weaker periodic structure in high-traffic pages.
-
SIGMA-ASL: Sensor-Integrated Multimodal Dataset for Sign Language Recognition
SIGMA-ASL is a multimodal dataset with 93,545 word-level ASL clips from Kinect RGB-D, mmWave radar, and dual IMUs, plus benchmarking protocols for single- and multi-modal recognition.
-
AegisTS: A Hierarchical Agent System with Reinforcement Learning for Multivariate Time Series Data Cleaning
AegisTS uses a two-level RL agent architecture with a dual-stage reward to jointly optimize cleaning order and method selection for multivariate time series, delivering up to 96% better cleaning quality and 27% better...
-
BadmintonGRF: A Multimodal Dataset and Benchmark for Markerless Ground Reaction Force Estimation in Badminton
BadmintonGRF is a new public multimodal dataset and benchmark that pairs multi-view video with instrumented GRF for markerless load estimation in badminton.
-
GAFSV-Net: A Vision Framework for Online Signature Verification
GAFSV-Net encodes online signatures as asymmetric Gramian Angular Field images and processes them with dual-branch ConvNeXt plus cross-attention to outperform sequence-based baselines on DeepSignDB and BiosecurID.
-
Autocorrelation Reintroduces Spectral Bias in KANs for Time Series Forecasting
Temporal autocorrelation reintroduces spectral bias in KANs for time series forecasting, which DCT preprocessing can mitigate.
-
A Convolutional Neural Network-Derived Catalog of Solar Flares from Soft X-Ray Observations
The CNN-derived catalog detects over seven times more solar flares than the GOES catalog and extends the power-law distribution of flare peak fluxes to smaller sizes.
-
Adversarial Robustness of Deep State Space Models for Forecasting
Spacetime SSM forecasters represent optimal Kalman predictors for autoregressive data but remain vulnerable to model-free attacks that exploit local linearity and increase error by over 33% compared to projected gradi...
-
CalM: A Self-Supervised Foundation Model for Population Dynamics in Calcium Imaging Data
CalM uses a discrete tokenizer and dual-axis autoregressive transformer pretrained self-supervised on calcium traces to outperform specialized baselines on population dynamics forecasting and adapt to superior behavio...
-
T1: One-to-One Channel-Head Binding for Multivariate Time-Series Imputation
T1 uses one-to-one channel-head binding in a CNN-Transformer hybrid to achieve robust multivariate time-series imputation, cutting average MSE by 46% versus the next-best baseline across 11 datasets even at 70% missingness.
-
MELT: A Behavioral Trace Dataset for High-Risk Memecoin Launch Detection
MELT is the first behavioral trace dataset for high-risk memecoin launch detection on Solana, providing 122 features, risk annotations, and ML benchmarks that reduce investment loss when used for selection.
-
Causal Time Series Generation via Diffusion Models
CaTSG is a unified diffusion model for causal time series generation that handles observational, interventional, and counterfactual tasks via backdoor adjustment and abduction-action-prediction.
-
Sundial: A Family of Highly Capable Time Series Foundation Models
Sundial uses TimeFlow Loss for native pre-training of Transformers on continuous time series from TimeBench, achieving SOTA point and probabilistic forecasting with millisecond inference.
-
What Causes Performance Degradation in Cross-Subject EEG Classification?
Controlled experiments attribute cross-subject EEG classification degradation to inter-subject variability in multi-class tasks and shortcut learning in single-class tasks.
-
Deep Time Series Models: A Comprehensive Survey and Benchmark
This survey and benchmark of deep time series models using the released TSLib library finds that models with specific structures perform well only on distinct analysis tasks.
-
Time-LLM: Time Series Forecasting by Reprogramming Large Language Models
Time-LLM reprograms frozen LLMs for time series forecasting via text prototypes and Prompt-as-Prefix, outperforming specialized models in standard, few-shot, and zero-shot settings.
-
Convolutional and Deep Learning based techniques for Time Series Ordinal Classification
First benchmarking of ordinal adaptations of CNN and DL methods for time series shows they outperform nominal TSC techniques on ordinal metrics across 29 selected problems.
-
A Time Series is Worth 64 Words: Long-term Forecasting with Transformers
PatchTST uses subseries patching and channel-independent Transformers to deliver significantly better long-term multivariate time series forecasting and strong self-supervised transfer performance.
-
Tropical time series, iterated-sums signatures and quasisymmetric functions
Defines iterated-sums signatures over commutative semirings (tropical case emphasized) for time-series feature extraction and links them to quasisymmetric functions over semirings.
-
Learning a Normal World Model for Few-Shot Boundary-Calibrated Abnormality Detection
A hypergraph entropic normal-world model learns normality from abundant normal sensor data and calibrates abnormality boundaries with few examples, reporting AUROC 0.9983 on the most complex NASA C-MAPSS subset.
-
INDEQS: Informed Neural controlled Differential EQuationS
INDEQS is a graph-informed NCDE variant that separates inner hidden-state mixing from outer vector-field mixing and reports lower MAE than uninformed NCDEs on synthetic advection data and real river/traffic tasks when...
-
SenFlow: Inter-Sentence Flow Modeling for AI-Generated Text Detection in Hybrid Documents
SenFlow integrates graph-based inter-sentence propagation with linear-chain CRF decoding for sentence-level AI-generated text detection and reports SOTA results on the new MOSAIC benchmark of 16k hybrid documents.
-
TS-Fault: Benchmarking Time Series Forecasters Against Structural Faults
TS-Fault benchmark finds clean-data accuracy anti-correlates with robustness to structural faults, with all catastrophic failures under mechanism-level faults and foundation models most fragile.
-
From Coarse to Fine: Managing Temporal Granularity in Spatio-Temporal Data for Fine-Grained Traffic Prediction
STRP is a granularity-aware model that predicts fine-grained spatio-temporal traffic from coarse inputs via tree convolution and inverse dilated convolution, outperforming baselines on six datasets in window-based and...
-
End-to-End Control of a Powered Knee-Ankle Prosthesis Towards Unified, Tuning-Free Assistance
An end-to-end TCN controller trained on multi-terrain data from 18 transfemoral amputees was deployed in real time on four users and reproduced key torque and angle scalings across five locomotion modes without subjec...
-
Second-Order Path Kernel Interpolation Formulas in Machine Learning
Derives second-order path-kernel interpolation formulas for gradient descent, SGD, and momentum training, adding curvature terms and a concentration estimate around the expected prediction.
-
Pretraining Recurrent Networks without Recurrence
SMT reduces RNN training to supervised learning on memory transitions (m_t, x_{t+1}) to m_{t+1} obtained from a Transformer encoder, enabling time-parallel training with O(1) gradient paths.
-
Dynamic Short Convolutions Improve Transformers
Dynamic short convolutions applied to key/query/value projections and linear layers in Transformers yield consistent performance gains and 1.33-1.60x compute advantages over standard models on language modeling from 1...
-
Dynamics Are Learned, Not Told: Semi-Supervised Discovery of Latent Dynamics Geometries For Zero-Shot Policy Adaptation
Contrastive learning bounds the Lipschitz constant of a trajectory dynamics encoder to support outcome-centric zero-shot adaptation in MuJoCo robotics tasks under severe dynamics shifts.
-
Beyond Holistic Models: Systematic Component-level Benchmarking of Deep Multivariate Time-Series Forecasting
TSCOMP is the first large-scale benchmark that deconstructs deep multivariate time series forecasters into fine-grained components, builds a corpus of over 20,000 evaluations, and shows that corpus-driven component se...
-
PHGNet: Prototype-Guided Hypergraph Construction for Heterogeneous Spatiotemporal Forecasting
PHGNet proposes prototype-guided hypergraph construction plus global-local representations and temporal attention to model high-order spatiotemporal dependencies in traffic data and reports better performance than pri...
-
ReactiveGWM: Steering NPC in Reactive Game World Models
ReactiveGWM introduces a decoupled diffusion architecture for player-NPC interactions that learns game-agnostic response logic for zero-shot strategy transfer across games.
-
Clin-JEPA: A Multi-Phase Co-Training Framework for Joint-Embedding Predictive Pretraining on EHR Patient Trajectories
Clin-JEPA is a multi-phase co-training framework for JEPA pretraining on EHR data that achieves convergent latent rollouts and improved multi-task AUROC on MIMIC-IV ICU records.
-
Clin-JEPA: A Multi-Phase Co-Training Framework for Joint-Embedding Predictive Pretraining on EHR Patient Trajectories
A five-phase co-training framework enables stable JEPA pretraining on EHR trajectories, producing converging latent rollouts and higher multi-task AUROC than baselines on MIMIC-IV ICU data.
-
What If We Let Forecasting Forget? A Sparse Bottleneck for Cross-Variable Dependencies
MS-FLOW uses a capacity-limited sparse routing mechanism to model only critical inter-variable dependencies in time series data, achieving state-of-the-art accuracy on 12 benchmarks with fewer but more reliable connections.
-
AegisTS: A Hierarchical Agent System with Reinforcement Learning for Multivariate Time Series Data Cleaning
AegisTS is a hierarchical agent system with RL that jointly optimizes cleaning order and method selection for MTS quality issues using a dual-stage reward based on cleaning and downstream performance.
-
Dynamics Aware Quadrupedal Locomotion via Intrinsic Dynamics Head
Concurrent training of an Intrinsic Dynamics Head with a dynamics reward yields more efficient and smoother quadrupedal locomotion policies that transfer to real robots with 12-18% gains in efficiency metrics.
-
Group Cognition Learning: Making Everything Better Through Governed Two-Stage Agents Collaboration
GCL uses a two-stage protocol with Routing, Auditing, Public-Factor, and Aggregation Agents to mitigate modality dominance and spurious coupling in multimodal learning, achieving state-of-the-art results on CMU-MOSI, ...
-
WISE-FM:Operation-Aware, Engineering-Informed Foundation Model for Multi-Task Well Design
WISE-FM is a design-aware, physics-informed multi-task foundation model that reduces virtual flow metering error by up to 13x on simulated wells and transfers to real Equinor data with high R-squared values by conditi...
-
Data-Driven Open-Loop Simulation for Digital-Twin Operator Decision Support in Wastewater Treatment
CCSS-RS achieves RMSE 0.696 and CRPS 0.349 at 1000-step horizons on a large public WWTP benchmark with 43% missingness, outperforming Neural CDE baselines by 40-46% in RMSE.
-
Conditional Attribution for Root Cause Analysis in Time-Series Anomaly Detection
Conditional attribution retrieves contextually similar normal states from VAE latent spaces and UMAP embeddings to explain time-series anomalies while preserving dependencies, improving root-cause accuracy on SWaT and...
-
VQ-Wave: A physics-driven spatio-temporal deep learning approach for non-contrast-enhanced lung ventilation and perfusion MRI
VQ-Wave is a physics-driven spatio-temporal neural network that learns to extract ventilation and perfusion maps from non-contrast lung MRI by training on simulated signals with amplitude modulations, frequency drifts...
-
Modern Structure-Aware Simplicial Spatiotemporal Neural Network
ModernSASST is the first simplicial complex-based spatiotemporal model that combines random walks on high-dimensional complexes with parallelizable temporal convolutional networks for efficient high-order topology capture.
-
AIBuildAI: An AI Agent for Automatically Building AI Models
AIBuildAI uses a manager agent and three LLM sub-agents to fully automate AI model development and achieves a 63.1% medal rate on MLE-Bench, matching experienced human engineers.
-
MISID: A Multimodal Multi-turn Dataset for Complex Intent Recognition in Strategic Deception Games
MISID is a multimodal multi-turn dataset for intent recognition in strategic deception games, paired with the FRACTAM framework that improves MLLM performance on hidden intent detection via decouple-anchor-reason steps.
-
BiTA: Bidirectional Gated Recurrent Unit-Transformer Aggregator in a Temporal Graph Network Framework for Alert Prediction in Computer Networks
BiTA redesigns temporal aggregation in TGNs by jointly using bidirectional GRU for sequential dependencies and Transformer for long-range context to improve alert prediction accuracy on real network data.
-
A General Framework for Generative Self-supervised Learning in Non-invasive Estimation of Physiological Parameters Using Photoplethysmography
TS2TC combines cross-temporal fusion generative anchor pretraining with dual-process transfer to achieve 2.49% lower RMSE than prior methods on PPG parameter estimation using only 10% labeled data.
-
ROMAN: A Multiscale Routing Operator for Convolutional Time Series Models
ROMAN converts time series into a shorter multiscale channel representation that lets standard CNN classifiers access scale and coarse-position information explicitly.
-
WiFlow: A Lightweight WiFi-based Continuous Human Pose Estimation Network with Spatio-Temporal Feature Decoupling
WiFlow achieves 97.25% PCK@20 and 99.48% PCK@50 on continuous pose estimation from WiFi CSI using a 2.23M-parameter network trained on 360,000 synchronized samples from 5 subjects.
-
Concurrence: A dependence criterion for time series, applied to biological data
Concurrence detects dependence between time series by training a classifier to separate aligned from misaligned segments.
-
X-IONet: Cross-Platform Inertial Odometry Network for Pedestrian and Legged Robot
X-IONet combines rule-based platform classification with a dual-stage attention network to predict displacement and uncertainty from IMU data, then fuses outputs via EKF, achieving reported error reductions on pedestr...
-
Vision-LLMs for Spatiotemporal Traffic Forecasting
ST-Vision-LLM reframes spatiotemporal traffic forecasting as vision-language fusion, using visual encoders on traffic grids and efficient numerical tokenization to achieve 15.6% better long-term accuracy and 30% gains...
-
Chinese Cyberbullying Detection: Dataset, Method, and Validation
Introduces CHNCI, the first Chinese cyberbullying incident detection dataset with 220,676 comments across 91 incidents, created via ensemble pseudo-labeling from explanation-generating methods followed by human annotation.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.