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..
FaCT: Faithful Concept Traces for Explaining Neural Network Decisions
Authors: Amin Parchami-Araghi, Sukrut Rao, Jonas Fischer, Bernt Schiele
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our models remain competitive on Image Net [12], while providing faithful concept-based explanations with a diverse (shared) concept basis. We quantitatively demonstrate that our concepts are more consistent, and through a user-study, that they are more interpretable than baselines. |
| Researcher Affiliation | Academia | Max Planck Institute for Informatics, Saarland Informatics Campus, Saarbrücken, Germany |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code available at github.com/m-parchami/Fa CT. |
| Open Datasets | Yes | Our models remain competitive on Image Net [12], while providing faithful concept-based explanations with a diverse (shared) concept basis. We quantitatively demonstrate that our concepts are more consistent, and through a user-study, that they are more interpretable than baselines. (and CUB [59] in Appendix K) datasets |
| Dataset Splits | Yes | Our SAEs are trained for reconstruction on Image Net [S30] s training set. We first take 50 samples per class as a held-out validation set and leave the rest for training. |
| Hardware Specification | No | Appendix A provides training details but does not specify exact GPU/CPU models, processor types, or memory used for experiments. Table I1 mentions inference times but not the hardware on which these measurements were made. |
| Software Dependencies | No | The paper mentions using pre-trained B-cos checkpoints and basing implementation on a Dictionary-Learning codebase, along with citing TorchVision. However, it does not provide specific version numbers for these software dependencies (e.g., PyTorch version, Python version, CUDA version). |
| Experiment Setup | Yes | For training the SAE, we use a batch size of 32,786 (individual feature vectors), with a sweep of learning rates λ [0.001, 0.0001], total number of latents K [8192, 16384], and sparsity factor of Top K-SAE topk [8, 16, 32] (except for Vi T @ Block 4/10, where we also tested topk = 64, as the lower values led to low accuracy). We use Adam Optimizer [S23], together with cosine learning-rate scheduler, with initial warm up of 2 epochs. We trained each model for 16 epochs |