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..
InfiGFusion: Graph-on-Logits Distillation via Efficient Gromov-Wasserstein for Model Fusion
Authors: Yuanyi Wang, Zhaoyi Yan, Yiming Zhang, Qi Zhou, Yanggan Gu, Fei Wu, Hongxia Yang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments across multiple fusion settings show that GLD consistently improves fusion quality and stability. Infi GFusion outperforms SOTA models and fusion baselines across 11 benchmarks spanning reasoning, coding, and mathematics. |
| Researcher Affiliation | Collaboration | 1The Hong Kong Polytechnic University, 2Infi X.ai, 3Zhejiang University EMAIL, EMAIL |
| Pseudocode | No | The paper describes algorithms and derivations but does not present them in a clearly labeled 'Pseudocode' or 'Algorithm' block format. Steps are described in paragraph text and mathematical equations. |
| Open Source Code | Yes | In section 4.1, we provide the data sources, and in the beginning of Appendix, we include an anonymous code repository URL for reproducibility. |
| Open Datasets | Yes | The questions are sourced from the Numina Math_1.52 dataset, while the answers are distilled from the Deep Seek-R1-671B model by the AM team3. Numina Math_1.5 represents the second iteration of the widely acclaimed Numina Math[37] dataset. It offers high-quality data for competition-level mathematical problems across various domains, including Algebra, Geometry, Combinatorics, Calculus, Inequalities, Logic and Puzzles, and Number Theory. (iii) In the code generation domain, we utilized the Kod Code-V1SFT-R1 dataset [38] |
| Dataset Splits | No | The paper mentions creating a multi-task training dataset comprising 130k examples, and specific sizes for general, mathematical, and coding datasets. However, it does not explicitly state the training, validation, or test splits used for these datasets within the main text or appendices. |
| Hardware Specification | Yes | Fusion uses a batch size of 16 on 8 80GB NVIDIA H800 GPUs for 20 hours. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies such as Python, PyTorch, or other libraries. It mentions 'C-Adam W' as an optimizer but without specific versioning for general software. |
| Experiment Setup | Yes | Fusion uses a batch size of 16 on 8 80GB NVIDIA H800 GPUs for 20 hours. We adopt early stopping at epoch 4 (of 5), learning rate 1 10 6, and default ULD weight λ = 0.5, GLD weight λ = 0.001. |