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..
Auto-Connect: Connectivity-Preserving RigFormer with Direct Preference Optimization
Authors: jingfeng Guo, Jian Liu, Jinnan Chen, Shiwei Mao, Changrong Hu, Puhua Jiang, Junlin Yu, Jing Xu, Qi Liu, LiXin Xu, Zhuo Chen, Chunchao Guo
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experiments on public benchmarks, we demonstrate that Auto-Connect substantially outperforms previous state-of-the-art methods, achieving superior results in joint location accuracy, topological consistency, and skinning quality. We summarize our contributions as follows: We introduce Auto-Connect, a novel automatic rigging pipeline that explicitly preserves skeletal connectivity through a connectivity-preserving tokenization scheme coupled with an enhanced pre-training framework, Rig Former. We develop a topology-aware reward function tailored for skeleton tree structures, and build upon this, we present a rigging post-training phase through our reward-guided DPO to further improve topology quality. To the best of our knowledge, this is the first work to combine reinforcement learning with the rigging task. We present a plug-and-play geodesic-aware bone probability prediction module that incorporates implicit geodesic features to dynamically determine the top-k bone for each vertex, effectively mitigating common skinning artifacts. |
| Researcher Affiliation | Collaboration | Jingfeng Guo1 , Jian Liu2,7 , Jinnan Chen3 , Shiwei Mao4,7, Changrong Hu5,7, Puhua Jiang4,7 Junlin Yu6,7, Jing Xu7, Qi Liu1 , Lixin Xu7, Zhuo Chen7, Chunchao Guo7 1South China University of Technology 2Hong Kong University of Science and Technology 3National University of Singapore 4Tsinghua Shenzhen International Graduate School 5University of Science and Technology of China 6Beijing Normal University 7Tencent Hunyuan |
| Pseudocode | No | The paper describes the methods in detailed text and illustrates concepts with figures such as Figure 1: Overview of the Auto-Connect and Figure 2: Connectivity-preserving tokenization process, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| 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 | Dataset. We evaluated our model on Articulation-XL2.0 (Art-XL2.0) [38] and Models Resource (MR) [33]. |
| Dataset Splits | Yes | Articulation-XL2.0 (Art-XL2.0) [38] and Models Resource (MR) [33]. Art-XL2.0 provides 46k samples for training and 2k samples for testing, while MR dataset contains 2.1k training samples and 540 testing samples. |
| Hardware Specification | Yes | Training is conducted on 8 H20 GPUs with a global batch size of 80, lasting 1 day for the Art-XL2.0 dataset and 10 hours for the MR dataset. |
| Software Dependencies | No | The paper mentions using the Adam optimizer and setting various hyperparameters but does not provide specific version numbers for programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow, CUDA). |
| Experiment Setup | Yes | The rigging pre-training stage uses a batch size of 192, lasting 2 days on the Art-XL2.0 dataset and 10 hours on the MR dataset, while the post-training stage runs 5 epochs on 14k curated preference pairs. For the skinning weight prediction stage, we set k = 6 follow the baseline. Training uses a batch size of 80, lasting 1 day for the Art-XL2.0 dataset and 10 hours for the MR dataset. We use the Adam optimizer with a base learning rate of 5 10 5, a weight decay of 0.001, and a linear warmup for the first 1,000 steps. For the DPO post-training stage, the optimizer remains unchanged, but the learning rate is reduced to 1 10 6, and the coefficient for LSFT is set to λ = 1. This stage performs 5 epochs on 14k curated preference pairs. The Rig Former model consists of 24 layers with a hidden dimension of 1024, and each transformer block incorporates a 16-head multi-head self-attention mechanism. |