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..
FRAM: Frobenius-Regularized Assignment Matching with Mixed-Precision Computing
Authors: Binrui Shen, LiangYuan, Shengxin Zhu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive CPU experiments show that FRAM consistently outperforms all baselines. On GPUs, with mixed precision, FRAM achieves up to a 370 speedup over its FP64 CPU implementation without sacrificing accuracy. |
| Researcher Affiliation | Academia | 1School of Mathematical Sciences, Laboratory of Mathematics and Complex Systems, MOE, Beijing Normal University, Beijing 100875, P.R. China 2Faculty of Arts and Sciences, Beijing Normal University, Zhuhai 519087, P.R. China 3School of Mathematical Sciences, Beijing Normal University, Beijing 100875, P.R. China 4Research Centers for Mathematics, Advanced Institute of Natural Science, Beijing Normal University, Zhuhai 519087, P.R. China 5Guangdong Provincial Key Laboratory of Interdisciplinary Research and Application for Data Science, BNU-HKBU United International College, Zhuhai 519087, P.R. China EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 Frobenius-Regularized Assignment Matching (FRAM) ... Algorithm 2 Scaling Doubly Stochastic Normalization (SDSN) |
| Open Source Code | Yes | 2https://github.com/Binrui Shen/FRAM |
| Open Datasets | Yes | The attributed graphs are constructed from a public dataset3, containing eight sets of pictures. The dataset covers five common transformations: viewpoint change, scale change, image blur, JPEG compression and illumination. The numerical results are presented in Table 2. 3http://www.robots.ox.ac.uk/~vgg/research/affine/ ... CMU house sequence4 is a classic benchmark dataset. It consists of a sequence of images showing a toy house captured from different viewpoints. Figure 3 demonstrates that FRAM achieves the best performance in both speed and accuracy on the House sequence dataset. 4https://www.cs.cmu.edu/afs/cs/project/vision/vasc/idb/images/motion/house/ ... The social network, comprising circles (or friends lists ) from Facebook [23], contains 4039 users (nodes) and 88234 relations (edges). We compare different methods in matching networks with noisy versions at 5%, 15% and 25%. Table 3 shows that FRAM achieves the highest node accuracy across all noise levels while maintaining computational efficiency. [23] Jure Leskovec and Andrej Krevl. SNAP Datasets: Stanford large network dataset collection. http://snap.stanford.edu/data, June 2014. |
| Dataset Splits | Yes | We compare different methods in matching networks with noisy versions at 5%, 15% and 25%... Real-world images. ... The running time and matching error are computed by averaging the results over five matching pairs (1 vs. 2, 2 vs. 3, ..., 5 vs. 6) from the same image set. House sequence ... Matching pairs consist of the first image and subsequent images with 5 image sequence gaps (such as image 1 vs. image 6 and so on). |
| Hardware Specification | Yes | All algorithmic comparison experiments are conducted in Python 3 on a workstation equipped with an Intel Core i7 (2.80 GHz) processor. ... For evaluating the mixed-precision design, we utilize a hardware platform equipped with an Intel Core i9-14900 (3.20 GHz) CPU and an NVIDIA RTX 4080 SUPER GPU. |
| Software Dependencies | No | All algorithmic comparison experiments are conducted in Python 3 on a workstation equipped with an Intel Core i7 (2.80 GHz) processor. |
| Experiment Setup | Yes | For FRAM, we set θ = 2 for attributed graph matching tasks and θ = 10 for unattributed tasks. Following [27], we set the regularization parameter to λ = 1 in our experiments, since the results are not sensitive to λ. We configure α to 0.95 to align with the parameter settings used in DSPFP [27]. |