Constrained Sequential Inference in Machine Learning Using Constraint Programming
Authors: Virasone Manibod, David Saikali, Gilles Pesant
IJCAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments in the context of molecule and music generation show that we can achieve the structure imposed post-training without straying too much from the structure of the dataset learned during training. |
| Researcher Affiliation | Collaboration | Virasone Manibod2 , David Saikali1 , Gilles Pesant1 1Department of Computer and Software Engineering, Polytechnique Montr eal, Montreal, Canada 2GIRO, Montreal, Canada EMAIL, EMAIL |
| Pseudocode | No | The paper describes its neurosymbolic framework Ge AI-BLAn C and its components, but does not provide any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | All our code and data are available4. [4https://github.com/cravethedave/Mini CPBP/tree/ijcai-2025] All our code and data are available6. [6https://github.com/Manibod/CMT CPBP] |
| Open Datasets | Yes | The GPT model, GPT2-ZINC480M-87M1 ... was trained on 480M molecules from the ZINC database2. [2https://zinc.docking.org/] The dataset used for the experiments is the EWLD [Simonetta et al., 2018] containing more than 5,000 lead sheets in many musical styles, but mainly Jazz, Pop and Rock. |
| Dataset Splits | Yes | After preprocessing, instances were split into training and test datasets of respective size 23,800 and 3,150. |
| Hardware Specification | Yes | Our experiments were run using an 8-core processor with a core speed of 4.20 GHz and 64GB of RAM. We ran all our experiments using 8 CPU cores, 48,000 M memory and a GPU type v100 with 32 Gi B memory. |
| Software Dependencies | No | The paper mentions specific software such as "Mini CPBP solver" and "Chord-conditioned Melody Transformer (CMT)" but does not provide specific version numbers for these tools as used in the experiments. |
| Experiment Setup | Yes | During generation we kept the model s default configuration with the following exceptions: we limited the generation to one new token at a time as per Fig. 1, we set the model s temperature to 1.5 which gives more varied results as reported by the model s authors, we decreased the maximum length of the molecule to fit our target length, and we disabled the early stopping parameter. For the CP part we use Mini CPBP3 with the default parameter values. After some initial experiments, we concluded that a geometric decay rk, with ratio r chosen so that the weight reaches about 0.7 once half the sequence has been generated (r = 0.9943)... As before we apply a geometric decay of the oracle constraint s weight, choosing ratio r = 0.985 so that the weight reaches around 0.4 once half the sequence has been generated. |