Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Dendritic Resonate-and-Fire Neuron for Effective and Efficient Long Sequence Modeling
Authors: Dehao Zhang, Malu Zhang, Shuai Wang, Jingya Wang, Wenjie Wei, Zeyu Ma, Guoqing Wang, Yang Yang, Haizhou Li
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our method maintains competitive accuracy while substantially ensuring sparse spikes without compromising computational efficiency during training. These results underscore its potential as an effective and efficient solution for long sequence modeling on edge platforms. |
| Researcher Affiliation | Academia | 1University of Electronic Science and Technology of China 2Shenzhen Loop Area Institute, 3The Chinese University of Hong Kong (Shenzhen) |
| Pseudocode | No | The paper describes the methodology using mathematical equations and textual explanations throughout sections 3 and 4, but does not include any clearly labeled 'Pseudocode' or 'Algorithm' blocks or figures. |
| Open Source Code | Yes | Justification: We mentioned our data in the Appendix. C.1 and code in supplemental material, respectively. |
| Open Datasets | Yes | To validate the effectiveness of our proposed method, we conduct experiments on multiple time-series datasets. All experiments are conducted at least five times. First, we compare the performance of D-RF with other SOTA models on commonly datasets, including Spiking Heidelberg Digits (SHD) [8] with 250 timesteps, Sequential MNIST, Permuted Sequential MNIST (S/PS-MNIST) [31] with 784 timesteps, and the more challenging Sequential CIFAR10 (S-CIFAR10) [7] with 1024 timesteps. As presented in Table 1, D-RF achieves SOTA performance while using fewer or comparable parameters. |
| Dataset Splits | Yes | Text [36]: A sentiment classification dataset based on the i MDB movie review dataset. The task is to classify a given movie review as positive or negative. The characters are encoded as integer tokens and the sequences are padded to a maximum length of 4,096. There are two classes representing positive and negative sentiment. The dataset includes 25,000 training samples and 25,000 test samples. Image [30]: An image classification dataset based on the CIFAR-10 dataset. It contains 32 32 grayscale images that are flattened into a sequence of length 1,024. The task is to classify each image into one of ten categories. The dataset contains 45,000 training samples, 5,000 validation samples, and 10,000 test samples. Pathfinder [32] A dataset for path finding tasks. It contains 32 32 grayscale images, each showing a start and end point represented as small circles. The task is to classify whether there is a dashed line (or path) connecting the start and end points. The sequences are padded to a length of 1,024. The dataset includes 160,000 training samples, 20,000 validation samples, and 20,000 test samples. |
| Hardware Specification | Yes | All experiments are conducted on Ubuntu server equipped with NVIDIA Ge Force RTX 4090 (24G Memory), Intel(R) Xeon(R) Platinum 8370C CPU@2.80GHz. |
| Software Dependencies | Yes | Py Torch 2.1.0, and CUDA 11.8. |
| Experiment Setup | Yes | The learning rate for neuron-specific parameters is set to 0.001, while a global learning rate of 0.005 is applied to the entire network. Detailed hyperparameter settings are summarized in Table 5. |