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..
TPP-SD: Accelerating Transformer Point Process Sampling with Speculative Decoding
Authors: Shukai Gong, YIYANG FU, Fengyuan Ran, Quyu Kong, Feng Zhou
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on both synthetic and real datasets demonstrate that our approach produces samples from identical distributions as standard methods, but with 2-6 speedup. Our ablation studies analyze the impact of hyperparameters such as draft length and draft model size on sampling efficiency. |
| Researcher Affiliation | Collaboration | 1Center for Applied Statistics and School of Statistics, Renmin University of China 2School of Information, Renmin University of China 3School of Cyber Science and Engineering, Wuhan University 4Alibaba Group |
| Pseudocode | Yes | Algorithm 1: TPP-SD |
| Open Source Code | Yes | Code is publicly available at https://github.com/GONGSHUKAI/tppsd. |
| Open Datasets | Yes | We consider three synthetic datasets: inhomogeneous Poisson, univariate Hawkes, and multivariate Hawkes processes... For real data, we consider four commonly used datasets: Taobao [1], Amazon [22], Taxi [33], and Stack Overflow [8]. |
| Dataset Splits | Yes | We consider three synthetic datasets: ... each with 1000 sequences within the time window [0, 100]. ... Details on data simulation procedures, data splitting, and experimental settings are provided in Appendices B.1 and C.3.1. Appendix B.1: each dataset contains 1000 sequences and is split into 80%/10%/10% for training/validation/testing. |
| Hardware Specification | Yes | All models were trained using the Adam optimizer [11] for up to 1000 epochs with a batch size of 16 on one single NVIDIA RTX 4090. |
| Software Dependencies | No | To implement our proposed CDF-based Transformer TPP model, we modified the codebase from [29]. The original RNN encoder for history aggregation was replaced with Transformer backbones proposed by THP [40], SAHP [36], and Att NHP [18]. ... All models were trained using the Adam optimizer [11]... |
| Experiment Setup | Yes | By default, we trained an 8-head, 20-layer target model and a 1-head, 1-layer draft model for each dataset. All models were trained using the Adam optimizer [11] for up to 1000 epochs with a batch size of 16 ... Early stopping based on validation log-likelihood was applied to prevent overfitting. In line with [29], we set the history embedding dimension D = 64 and the number of mixture components M = 64. |