EnCoBo, a post-concept bottleneck model for generative models, addresses the problems of poor interpretability and interrogation by restricting representations to flow solely through explicit concepts without auxiliary visual cues. Unlike autoencoder-based approaches, it utilizes a decoder-less, energy-based framework to guide generation directly in the latent space. Guided by a diffusion-scheduled energy function, EnCoBo supports powerful post-concept interrogation, such as concept formation and negation, across arbitrary concepts. Experimental results on the CelebA-HQ and CUB datasets demonstrate that EnCoBo improves concept-level human interrogation and interpretability while maintaining competitive visual quality.
