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..
Mesh-RFT: Enhancing Mesh Generation via Fine-grained Reinforcement Fine-Tuning
Authors: Jian Liu, Jing Xu, Song Guo, Jing Li, jingfeng Guo, Jiaao Yu, Haohan Weng, Biwen Lei, Xianghui Yang, Zhuo Chen, Fangqi Zhu, Tao Han, Chunchao Guo
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiment results show that our M-DPO approach reduces Hausdorff Distance (HD) by 24.6% and improves Topology Score (TS) by 3.8% over pre-trained models, while outperforming global DPO methods with a 17.4% HD reduction and 4.9% TS gain. These results demonstrate Mesh-RFT s ability to improve geometric integrity and topological regularity, achieving new state-of-the-art performance in productionready mesh generation. Extensive experiments across diverse meshes demonstrate Mesh-RFT s superior performance, achieving significant improvements over both the pretrain baseline (24.6% HD reduction, 3.8% TS improvement) and global DPO (17.4% HD reduction, 4.9% TS improvement), establishing a new benchmark for accuracy and fidelity in generative mesh modeling. 4 Experiments 4.1 Experiment Settings 4.2 Qualitative Results 4.3 Quantitative Results 4.4 Ablation Study |
| Researcher Affiliation | Collaboration | 1 Hong Kong University of Science and Technology 2 Tencent Hunyuan 3 University of Science and Technology of China 4 South China University of Technology |
| Pseudocode | No | The paper describes the Mesh-RFT framework and its components (Mesh Generation Pre-training, Preference Dataset Construction, Masked Direct Preference Optimization) using textual descriptions and mathematical formulas (e.g., LM-DPO objective, metric definitions). However, there are no explicitly labeled pseudocode or algorithm blocks presenting structured steps of the methods. |
| Open Source Code | No | While we currently do not provide open access to the data and code, we plan to release the code along with sufficient instructions to reproduce the main experimental results after the paper has been accepted. |
| Open Datasets | Yes | Our model is pretrained on 2M meshes from large-scale datasets including Shape Net V2 [63], 3D-FUTURE [64], Objaverse [65], Objaverse-XL [66], and licensed assets. |
| Dataset Splits | Yes | After filtering low-quality scans and poorly topologized CAD models, 800K meshes form the fine-tuning subset. For preference alignment, we construct a specialized dataset of 10,000 generated meshes, each paired with 8 topological variations derived from the same input point cloud. To enhance geometric generalization, meshes are perturbed at the vertex level and subsampled from an initial 50K-point cloud to 16,384 points, without enforcing watertightness. For evaluation, we employ two test sets: (1) 100 high-quality, artist-designed meshes for qualitative analysis, and (2) 100 dense, out-of-distribution meshes generated by Hunyuan2.5 [48], providing rigorous real-world validation. |
| Hardware Specification | Yes | We pretrained on 256 NVIDIA H20 GPUs (2/GPU) for 10 days with Adam W [67] (β1 = 0.9, β2 = 0.99) and Flash Attention, following a 100-step linear warm-up. |
| Software Dependencies | No | The paper mentions techniques and optimizers like 'Adam W [67]' and 'Flash Attention', but it does not specify version numbers for any software dependencies, libraries, or frameworks (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | We pretrained on 256 NVIDIA H20 GPUs (2/GPU) for 10 days with Adam W [67] (β1 = 0.9, β2 = 0.99) and Flash Attention, following a 100-step linear warm-up. M-DPO post-training took 8 hours on 64 GPUs with a 5e 7 learning rate. Our model consists of 24 Transformer layers (1.1B parameters) arranged in a three-stage hourglass structure(2-4-12-4-2). It features a hidden dimension of 1536 and 16 attention heads. The vocabulary size for vertex coordinate quantization is 1024. The architecture supports a 36,864token context window during inference and generates meshes through temperature-controlled sampling (T = 0.5), balancing output diversity and stability. |