CFM: Language-aligned Concept Foundation Model for Vision

AI summary

Overview

Research area: Interpretable computer vision and vision-language models (concept-based interpretability, sparse autoencoders, vision foundation models).

Technical level: Intermediate — readers should be comfortable with CLIP-style vision-language models, sparse autoencoders, and standard segmentation/classification benchmarks, but the paper's core ideas are described conceptually.

Scope: The paper introduces CFM, a language-aligned concept foundation model that produces spatially grounded, hierarchically organized, and automatically named concepts usable as explanations for any downstream task of a CLIP-based vision encoder.

What This Paper Is About

Language-aligned vision foundation models like CLIP perform strongly across many tasks, but their internal representations are opaque, so it is hard to know why they make a decision. Existing methods try to decompose these representations into human-interpretable concepts, but they suffer from poor spatial grounding in the image, flat (non-hierarchical) concept sets, limited naming vocabularies, and they have only been used for global tasks such as image classification. CFM addresses these gaps by learning fine-grained concepts that are localized to image patches, organized into explicit parent-child hierarchies, and given names using a hierarchy-aware matching scheme — while remaining competitive with opaque models on classification, segmentation, and captioning.

Key Contributions

1. A concept foundation model (CFM) that lifts concept bottleneck models to the scale of language-aligned vision foundation models, using unified local concept representations applicable to any downstream task, not just image classification.
2. Explicit hierarchical relations between concepts, discovered from patch-wise co-occurrence of concept activations, rather than imposed from a predefined taxonomy.
3. An improved, hierarchy-aware concept labeling scheme that offsets child concept dictionary vectors with those of their parents and uses an architecture-aware text encoding.
4. Empirical validation that CFM produces concepts that are more localized, consistent, pure, hierarchically organized, and meaningfully named than prior work — without degrading downstream task performance, supported by quantitative metrics and a human user study.

Main Findings

• Concepts are better localized, more consistent, and purer. Using metrics adapted from Pham et al. on PartImageNet and COCO-Stuff, CFM scored 44.3 (Part) and 44.5 (Stuff) for locality, 15.3 and 13.1 for consistency, and 0.333 and 0.539 for impurity, compared with PatchSAE (32.2, 34.6, 13.1, 12.7, 0.353, 0.663) and SALF-CBM (16.4, 19.5, 9.9, 6.7, 0.841, 1.143). CFM also had the highest annotation-free C²-Score of 0.465 (vs 0.298 for SALF-CBM and 0.295 for PatchSAE), computed over ImageNet in DINOv2 feature space.
• Humans judge the concepts favorably. In an Amazon MTurk study covering roughly 1000 neurons with 5 users each, users rated concepts mostly or completely consistent for more than half of the shown concepts, and more than 80% of annotators assigned at least "somewhat consistent" (score 3 or higher).
• Discovered hierarchies are directionally accurate. A gpt-5.5 model judge asked to evaluate parent-child direction from top-5 images and their activating regions obtained 79.79% accuracy on pairs found by CFM.
• Concept names align with human labels. A user study comparing annotator-provided 1-2 word descriptions against CFM's assigned names using SentenceBERT cosine similarity found a statistically significant difference (p < 2.22 × 10⁻¹⁶, Wilcoxon rank-sum test) over randomly drawn names. A gpt-5-mini judge evaluating whether top-10 concepts are present in COCO validation images found CFM's names most accurate among compared methods and with the biggest improvement over the random baseline.
• Classification is competitive while remaining grounded. On ImageNet and Places365, CFM reached 78.9 and 55.4, compared with DN-CBM (79.5, 55.1), CDM (79.3, 52.6), CF-CBM (78.5), LaBo (78.9), LF-CBM (75.4, 50.6), SALF-CBM (76.3, 49.4), D-CBM (70.5, 50.9), DCLIP (68.0, 40.3), zero-shot CLIP (68.6, 41.2), and a linear probe (80.2, 55.1). The paper notes all baselines except SALF-CBM and CFM provide only image-level concepts, and that baselines were trained explicitly and only for classification.
• Captioning performance is preserved and steerable. Against its opaque backbone (CLIP-DINO), CFM scored 0.72 vs 0.72 on CLIP-Sc, 0.72 vs 0.73 on RefCLIP-Sc, and 0.35 vs 0.35 on Capture. Because concepts are named, the captioning model can be steered at the concept level (for example, removing a "cake" concept while preserving a "drink" concept).
• Open-vocabulary segmentation is on par with the strongest opaque baselines. CFM averaged 41.3 mIoU across benchmarks and 41.9 with AnyUp, versus CLIP-DINOiser at 41.4 and 42.2 with AnyUp; per-dataset CFM scored 80.7 (VOC20), 36.5 (C59), 24.2 (Stuff), 38.5 (City), 20.7 (ADE), 33.1 (Context), 34.7 (Object), and 62.2 (VOC). The paper states CFM is the only method in the comparison to provide explanations for its predicted segmentation.
• Larger dictionaries trade interpretability against classification accuracy. Ablating 4k/8k/16k, ImageNet accuracy rose 78.2 → 78.6 → 78.6 and Places365 55.5 → 55.6 → 56.0, while the C² score fell 0.488 → 0.465 → 0.418. Average OVS was 41.1 / 41.3 / 40.8. The authors selected 8k as the default to balance segmentation performance with computational efficiency.

Methodology in Plain English

The authors start from a pretrained CLIP vision encoder and work at the level of image patches rather than whole images. Because raw CLIP patch tokens are noisy, they first apply a "DINOising" step inspired by CLIP-DINOiser: matching self-supervised DINO features are used like a voting system that smooths representations of semantically similar patches and attenuates noise, approximated for efficiency by a small 3×3 convolutional layer trained on patch tokens.

On top of these smoothed patch features they train a sparse autoencoder — a lightweight encoder and decoder that map each patch into a high-dimensional, sparse space where each dimension ideally corresponds to one interpretable concept, trained with a reconstruction loss and a sparsity constraint. They use a BatchTopK sparsity objective plus a Matryoshka objective, which trains nested prefixes of the concept vector to each reconstruct the full features; this makes early dimensions capture coarse, high-level information and later dimensions capture residual fine-grained detail.

Spatial grounding comes for free from encoding each patch: a concept's activation across patches forms a heat map over the image. To recover relationships, the authors build a co-occurrence matrix of concept activations across patches — the weighted conditional probability that concept i is active given that concept j is — and treat i as a parent of j when this probability exceeds a threshold τ. Finally, concepts are named by matching their decoder vectors to CLIP text embeddings of a large vocabulary; for fine-grained concepts, the child's decoder vector is offset by its discovered parents' vectors plus the decoder bias before matching, which the authors found is necessary for accurate names.

They validate with interpretability metrics (locality, consistency, impurity, C²-Score), human studies on Amazon MTurk, and automated LLM judges, and then test the same concept representation on classification, image captioning (via an MLP adapter into a Gemma-2-2B Instruct language model trained on CC12M DreamLIP Long captions), and open-vocabulary segmentation.

Why This Matters

Impact on research: The paper argues that concept-based interpretability has been stuck at the level of global classification, and that a concept representation must be local, grounded, and hierarchical to serve as a foundation model component. By demonstrating that interpretability can be added without sacrificing accuracy on classification, segmentation, and captioning, it reframes explanations as a general-purpose property of the representation rather than a task-specific add-on. The authors claim CFM is the first concept-based foundation model unlocking explanations for all of its downstream vision tasks, and they release code and interactive visualizations.

Real-world applications:

• Safety-critical visual systems (e.g., medical or industrial inspection) where users need to know whether a model is right for the right reasons.
• Debugging and auditing deployed vision models, including investigating why a prediction failed and whether it relied on spurious cues.
• Controllable generation and captioning, where users can suppress or enforce specific concepts (such as removing "cake" while keeping "drink").
• Open-vocabulary segmentation for domains without fixed label sets, where each predicted region can be explained by a linear combination of named concepts.

Industry relevance: Because CFM is built on frozen CLIP backbones and adds a lightweight sparse autoencoder, it fits the increasingly common pattern of interpreting an existing foundation model rather than retraining one. The ability to steer downstream outputs through named concepts is directly useful for content moderation, compliance, and human-in-the-loop editing workflows.

Future Directions

• Improving concept naming, which the authors explicitly identify as still having room for improvement despite gains over prior work.
• Developing evaluation strategies for explanations that suit concepts learned in an unsupervised manner, since most existing explanation metrics are not directly applicable.
• Learning the concept representation on larger, more diverse web-scale datasets to build a broader concept space.
• Extending interpretable vision tokens to large vision-language models for tasks such as visual question answering.
• The authors also report structural ablations (scaling to the 400M-parameter SigLIP-2 backbone, the necessity of guided pooling, varying the sparsity factor k, and comparisons against alternative SAE objectives) in Appendix B.8, which point to further architectural exploration.

Target Audience

This paper is most valuable to interpretability and explainable-AI researchers, practitioners building on CLIP-family vision foundation models, and engineers who need auditable or steerable vision systems in domains such as safety-critical inspection, medical imaging, and content moderation. It is also useful for readers of concept bottleneck model literature who want to understand how those ideas extend to local, spatially grounded, and hierarchical representations. Some familiarity with sparse autoencoders, CLIP embedding spaces, and standard segmentation benchmarks is assumed.

Authors’ abstract

Language-aligned vision foundation models perform strongly across diverse downstream tasks. Yet, their learned representations remain opaque, making interpreting their decision-making difficult. Recent work decompose these representations into human-interpretable concepts, but provide poor spatial grounding and are limited to image classification tasks. In this work, we propose CFM, a language-aligned concept foundation model for vision that provides fine-grained concepts, which are human-interpretable and spatially grounded in the input image. When paired with a foundation model with strong semantic representations, we get explanations for any of its downstream tasks. Examining local co-occurrence dependencies of concepts allows us to define concept relationships through which we improve concept naming and obtain richer explanations. On benchmark data, we show that CFM provides performance on classification, segmentation, and captioning that is competitive with opaque foundation models while providing fine-grained, high quality concept-based explanations. Code at https://github.com/kawi19/CFM. Interactive visualizations at https://concept-foundation-model.mpi-inf.mpg.de.