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..
PASS: Path-selective State Space Model for Event-based Recognition
Authors: Jiazhou Zhou, Kanghao Chen, Lei Zhang, Lin Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments across five public and three proposed datasets demonstrate PASS s superior performance. It outperforms previous methods on five public datasets and shows strong generalization across varying inference frequencies with less accuracy drop (ours -8.62% v.s. -20.69% for the baseline). |
| Researcher Affiliation | Academia | Jiazhou Zhou AI Thrust, HKUST(GZ) International Digital Economy Academy EMAIL; Kanghao Chen AI Thrust, HKUST(GZ) EMAIL; Lei Zhang International Digital Economy Academy EMAIL; Lin Wang School of Electrical and Electronic Engineering, Nanyang Technological University EMAIL |
| Pseudocode | Yes | A.2 Py Torch-style Pseudocode for PEAS Module In Algorithm 1, we present the Py Torch-style pseudocode of the proposed PEAS module to facilitate readers understanding. |
| Open Source Code | Yes | Project page: https://github.com/jiazhou-garland/PASS_Homepage. Corresponding Author. |
| Open Datasets | Yes | Five publicly available event datasets are evaluated in this paper, including PAF [41], Se Act [77], HARDVS [63], N-Image Net [29] and N-Caltech101 [43]. Our Ar DVS100, Real-Ar DVS10 and Tem Ar DVS100 Dataset. Existing datasets only provide events within a limited distribution of event length (106 for objects and 107 for actions). We introduce the Ar DVS100 and Tem Ar DVS across a broad distribution of event length (106 to 109), synthesized by concatenating event streams with varying meta actions, thus capturing action transitions over time. |
| Dataset Splits | Yes | We allocated 80% for training and 20% for testing (evaluating). Additionally, to assess the model s real-world applicability, we created a real-world dataset, named Real-Ar DVS10, comprising event-based actions lasting from 2s to 75s, encompassing 10 distinct classes selected from the Ar DVS100 datasets. The train and validation (test) split ratio is 7:3. |
| Hardware Specification | No | 8. Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] . Justification: The paper details the computational resources in the appendix. |
| Software Dependencies | No | A.2 Py Torch-style Pseudocode for PEAS Module In Algorithm 1, we present the Py Torch-style pseudocode of the proposed PEAS module to facilitate readers understanding. |
| Experiment Setup | Yes | We utilize the default hyperparameters for the B-Mamba layer [78], setting the state dimension to 16 and the expansion ratio to 2. Additionally, we adjust the stochastic depth ratio to 0, 0.15, and 0.5 for the Tiny, Small, and Middle versions, respectively. We utilize the Adam W optimizer with a cosine learning rate schedule with the initial 5 epochs for linear warm-up. Unless a special statement is made, the default settings for the learning rate and weight decay are 1e-3 and 0.05, respectively. The model is trained with 100 epochs for PAF, Se Act, and N-Caltech101 datasets and 50 epochs for HARDVS, N-Imagenet, Ar DVS100, Tem Ar DVS100, and Real-Ar DVS10 datasets. We employ BFloat16 precision during training to improve stability. For data augmentation, we implement random scaling, random cropping, random flipping, and data mixup of the event frames during the training phase. |