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..
Curriculum Abductive Learning
Authors: Wen-Chao Hu, Qi-Jie Li, Lin-Han Jia, Cunjing Ge, Yu-Feng Li, Yuan Jiang, Zhi-Hua Zhou
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments across multiple tasks show that C-ABL outperforms previous ABL implementations, significantly improves training stability, convergence speed, and final accuracy, especially under complex knowledge setting. [...] Experiments on digit addition, chess attack, and real-world legal judgment tasks demonstrate that C-ABL significantly improves training stability, convergence speed, and final accuracy over strong ABL and neuro-symbolic baselines. |
| Researcher Affiliation | Academia | 1National Key Laboratory for Novel Software Technology, Nanjing University, China 2School of Artificial Intelligence, Nanjing University, China EMAIL,EMAIL |
| Pseudocode | Yes | The pseudocode is provided in Algorithm 2 in Appendix D. [...] Algorithm 2: Curriculum Abductive Learning (C-ABL) Training |
| Open Source Code | Yes | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: We have provided code in the supplementary material. |
| Open Datasets | Yes | We use images from CIFAR [Krizhevsky et al., 2009] to represent digits. [...] constructed using the first 16 classes of CIFAR100 [Krizhevsky et al., 2009]. [...] In the decimal setting, we now use digit images from MNIST dataset [Krizhevsky et al., 2009], and in the hexadecimal setting, we use digits 0 9 and letters A F from EMNIST dataset [Cohen et al., 2017]. [...] We use the dataset from Huang et al. [2020], which includes 687 records. |
| Dataset Splits | No | In phase p, training is guided by sub-base KBp. To align the model s prediction space with the reasoning scope of KBp, we dynamically schedule training data whose concept labels fall within the domain Zp [Bengio et al., 2009, Marconato et al., 2023a]. Then, each phase follows the standard ABL procedure, as stated in Section 2.3. [...] We compare C-ABL with prior ABL implementations under two semi-supervised training settings, where 10% and 50% of the data include ground-truth concept labels available for training supervision. |
| Hardware Specification | Yes | All experiments are performed on a server with Intel Xeon Gold 6226R CPU and Tesla A100 GPU, and each experiment is repeated 5 times. |
| Software Dependencies | No | The paper mentions using ResNet18 [He et al., 2016], LeNet-5 [Le Cun et al., 1998] and google-bert/bert-base-chinese [Devlin et al., 2019] as perception/learning modules, and refers to "Ablkit: a python toolkit for abductive learning" [Huang et al., 2024]. However, specific version numbers for underlying software dependencies (e.g., Python, PyTorch, TensorFlow) are not explicitly provided in the text. |
| Experiment Setup | Yes | All methods use Res Net18 [He et al., 2016] as the perception module, and are trained for a total of 5,000 iterations. [...] All methods use Le Net-5 [Le Cun et al., 1998] as the perception model, and are trained for a total of 1,000 iterations. [...] We use the pretrained google-bert/bert-base-chinese [Devlin et al., 2019] as the learning model. [...] We set the minimum phase size to τ = 2. |