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..
Object-Centric Concept-Bottlenecks
Authors: David Steinmann, Wolfgang Stammer, Antonia Wüst, Kristian Kersting
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate OCB on complex image datasets and conduct a comprehensive ablation study to analyze key components of the framework, such as strategies for aggregating object-concept encodings. The results show that OCB outperforms traditional CBMs and allows one to make interpretable decisions for complex visual tasks. |
| Researcher Affiliation | Academia | 1Computer Science Department, TU Darmstadt; 2Hessian Center for AI (hessian.AI); 3German Research Center for AI (DFKI); 4Centre for Cognitive Science, TU Darmstadt |
| Pseudocode | Yes | cf. full filtering pseudocode in Alg. 1 in the appendix. (Appendix B: Algorithm 1 Refine Object Proposals) |
| Open Source Code | Yes | Code and data available at: https://github.com/DavSte13/Object-Centric-Concept-Bottlenecks/ |
| Open Datasets | Yes | We conduct experiments across multiple datasets, including our newly introduced COCOLogic benchmark, based on MSCOCO (Lin et al., 2014)... For multi-label settings, we use PASCAL-VOC (Everingham et al., 2010) and MSCOCO (Lin et al., 2014)... For single-label classification, we use SUN397 (Xiao et al., 2010, 2016) and our novel COCOLogic dataset (see Sec. 4). |
| Dataset Splits | Yes | The COCOLogic dataset comprises ten semantically rich classes derived from COCO images. ...Each class is defined by a specific logical rule, detailed in Tab. 3 alongside the number of training and test examples per class. ...For its training, we optimized the parameters learning rate and the number of epochs the layer was trained for, using the validation sets. |
| Hardware Specification | Yes | All experiments were conducted using T single GPUs from Nvidia DGX2 machines equipped with A100-40G and A100-80G graphics processing units. |
| Software Dependencies | No | The paper mentions using Sp Li CE and CLIP but does not provide specific version numbers for these or any other software libraries or frameworks used. |
| Experiment Setup | Yes | We followed their default settings, for model, vocabulary and l1-regularization. The only exception is to that is CBM (equal capacity), where we reduced the regularization from 0.25 to 0.2, whcih led to a comparable number of non-zero concepts to OCB. For the predictor network, we used a single linear layer. For its training, we optimized the parameters learning rate and the number of epochs the layer was trained for, using the validation sets. The parameter configurations for each dataset and aggregation can be found in the code repository. |