Multimodal Concept Bottleneck Models

Tongqing Shi, Ge Yan, Tuomas Oikarinen, Tsui-Wei (Lily) Weng

University of California San Diego

Preprint 2025

Abstract

Concept Bottleneck Models (CBMs) enhance the interpretability of deep neural networks by aligning the features extracted from images with human-interpretable concepts. However, existing CBMs rely on linear classifiers, which not only restricts them to a fixed set of predefined classes, but also allows models to bypass concept activations and compensate with non-concept predictive cues.

In this paper, we propose Multimodal Concept Bottleneck Model (MM-CBM) to address these issues and extend CBMs to the CLIP framework.

MM-CBM employs dual concept bottleneck layers to align both image and text embeddings into a shared interpretable concept space.
This design enables interpretable reasoning for tasks such as zero-shot classification and image retrieval.
MM-CBM achieves improved accuracy across four benchmarks, while remaining comparable to black-box model, demonstrating a strong balance between interpretability and performance.

Fig 1: Inference process comparison of MMCBM with conventional CBMs.

Method

MMCBM introduces a multimodal extension of Concept Bottleneck Models by integrating both visual and textual modalities into a unified concept-driven reasoning pipeline.

Fig 2: Overview of MM-CBM. (Left) A. Extracting high-quality concept annotations for each modality. (Right) B. Using an auxiliary dataset to train dual CBLs, jointly optimizing interpretability alignment and task performance.

Step A: Collecting concept activation data

We construct a multimodal concept activation dataset by aligning image–text inputs with a shared concept space using pretrained object-detection and language models. Step A consists of the following steps:

Generate candidate concepts using LLMs following prior work[2].
Extract visual concept signals via Grounding DINO Swin-B, converting detected regions into one-hot concept labels.
Compute textual concept relevance using all-mpnet-base-v2 through semantic similarity.

This process generates large-scale supervision labels that reflect how visual and textual inputs relate to semantic concepts.

Step B: Training dual concept bottleneck layers

In step B, we jointly train dual concept bottlenecks for images and texts to align concept representations and enable interpretable reasoning in a unified concept space. Through concept supervision and cross-modal alignment objectives, the model learns to perform reasoning explicitly in the concept space, enabling interpretable and transparent predictions. We introduce three strategies to enhance the interpretability of our multimodal CBM model.

Generation of rich textual information - We utilize vision-language models to generate detailed textual descriptions for images, mitigating the label scarcity problem commonly observed in conventional image datasets.
Number of effective concepts (NEC) - Following previous work[4], we enforce sparsity by restricting the final decision to rely on only a few salient concept activations, thereby enhancing interpretability.
Non-negative concept representation space - To eliminate the impact of irrelevant or negatively correlated concepts, we apply a masking strategy that sets uninformative concept activations to zero on a per-sample basis.

Experimental Results

1. Task performance: compared with other CBMs

MM-CBM achieves performance comparable to the strongest baseline VLG-CBM, and surpasses others by over 10% accuracy on ImageNet.

Table 1:Comparison with LF-CBM[2], LM4CV[10], LaBo[11], VLG-CBM[4] on ANEC-5 using CLIP RN50. Best results for each benchmark are in bold; second-best are underlined.

2. Task performance: compared with black-box CLIP

Across seven datasets, MM-CBM attains performance comparable to CLIP’s linear-probe and zero-shot results.

Table 2: Test accuracy comparison with black-box CLIP[1] on zero-shot and finetuned settings.

3. Interpretability results

Human study: We compared two variants of MM-CBM (zero-shot and fine-tuned) against SpLiCE[3]. For each method, participants were shown 100 randomly sampled ImageNet images with the correct label. For every sample, we displayed the top five contributing concepts and their corresponding weights, and asked users to rate which explanation is better? on a Likert scale from 1 to 5.

Fig 3: Interpretability results compared with SpLiCE[3].

VLM judge: We prompted VLM with the template: "Why is this image categorized as {cls}?", and collected 5,000 explanation pairs from imagenet dataset. To evaluate which explanation contained more informative visual concepts, we prompt another VLM by asking: "Which description has more informative visual concepts in this image?". In our experiments, we used LLaVA-v1.5-7B and Llama-3.2-11B-Vision-Instruct, alternating their roles to ensure fairness and robustness. MM-CBM explanations were preferred over LLaVA 1.5 in 4,433 (88.7%) out of 5,000 cases, and over Llama 3.2 in 3,292 (65.8%) cases.

Fig 4: Visualization of Interpretability results.