concept bottleneck modeling

This work addresses the limitations of existing concept-based models, which rely on fine-grained annotations and treat concepts as flat, independent units, thereby hindering the construction of interpretable, hierarchical concept structures. To overcome this, the authors propose Multi-Level Concept Segmentation (MLCS) and Deep Hierarchical Concept Embedding Models (Deep-HiCEMs), which require only coarse-grained top-level supervision to automatically discover multi-layered, human-interpretable concept hierarchies. The framework supports concept interventions across abstraction levels and successfully uncovers novel, explainable concepts absent from training data across multiple benchmarks. While maintaining high predictive accuracy, the method significantly enhances task performance through test-time interventions, marking the first approach capable of automatically constructing a multi-granular concept system from coarse-grained labels alone.

Existing concept embedding methods struggle to model inter-concept relationships and rely heavily on multi-granularity human annotations, limiting both interpretability and practical applicability. This work proposes Hierarchical Concept Embedding Models (HiCEMs), which introduce hierarchical structure into concept representations for the first time and integrate an unsupervised concept splitting technique to automatically discover fine-grained subconcepts from pretrained models. HiCEMs generate multi-level, interpretable embeddings without requiring additional annotations and support test-time multi-granularity interventions. Evaluated across multiple datasets—including the newly introduced PseudoKitchens—the approach demonstrates its ability to uncover human-understandable subconcepts, improve task accuracy, and enable effective explanation and intervention.

Weak interpretability of large language models (LLMs), unreliability of post-hoc explanation methods, and limitations of concept bottleneck models (CBMs)—including dependence on costly human annotations, restricted representational capacity, and lack of interpretability at the task level—motivate this work. We propose the self-supervised Interpretable Concept Embedding Model (ICEM), the first framework to introduce concept modeling into the textual domain. ICEM leverages the inherent generalization capability of LLMs to autonomously predict concept labels without manual annotation. It enables end-to-end interpretable prediction via concept embeddings and an interpretable decision function, supporting concept intervention, logical attribution, and controllable decoding-path steering. On text classification tasks, ICEM achieves performance comparable to fully supervised CBMs and black-box LLMs, while providing human-understandable, causally grounded explanations. Thus, ICEM unifies interpretability, interactivity, and controllability in a single architecture.

This work addresses the fundamental trade-off in machine learning between high predictive performance and low interpretability inherent in “black-box” models (e.g., deep neural networks, ensemble methods). It rigorously distinguishes post-hoc explanation—applied after model training—from inherently interpretable modeling—designed for transparency from inception. To reconcile accuracy and interpretability, we propose a hybrid modeling paradigm centered on symbolic knowledge embedding, integrating differentiable symbolic modules, knowledge distillation, and symbolic reasoning into the model architecture itself. This enables joint optimization of fidelity and interpretability at the design stage. Extensive experiments across diverse domains demonstrate that our approach matches the predictive accuracy of state-of-the-art black-box models while generating human-understandable, logically grounded decision rules. As a result, it substantially enhances model trustworthiness and deployment viability in safety- and accountability-critical applications.

The causal origins of interpretable units—such as induction heads—in large language models remain poorly understood. This work proposes a scalable mechanistic data attribution framework that integrates influence functions with causal interventions to establish, for the first time, direct causal links between specific training examples and the emergence of such interpretable components. The study reveals that structured repetitive data plays a catalytic role in circuit formation and demonstrates a direct functional relationship between induction heads and in-context learning capabilities. By selectively intervening on a small set of high-influence training samples, the emergence of attention heads can be significantly modulated. Furthermore, the proposed data augmentation strategy consistently accelerates circuit convergence across different model scales.

Conventional concept-based interpretability methods provide only local explanations, failing to capture global neural circuit dynamics in deep models. Method: We propose the first concept-driven framework for global neural circuit analysis, introducing a model-agnostic, hierarchical concept analysis pipeline that quantifies the emergence, interaction, and propagation of semantic concepts across layers; we further design BAGEL—a visualization platform that encodes concept–class relationships as structured knowledge graphs to diagnose spurious correlations and data biases. Contributions/Results: (1) First systematic, end-to-end tracking of high-level semantic concepts throughout deep model internals; (2) Discovery of decision-critical latent neural circuits and information flow patterns; (3) Effective identification of generalization failures induced by data bias, substantially improving explanation fidelity and debugging capability.

Interpretability of neural network internal activations has long been constrained by hand-crafted assumptions and scalability limitations of surrogate models. Method: We propose the first end-to-end trainable interpretability assistant that frames interpretability as a prediction task: a sparse concept encoder—acting as a communication bottleneck—maps internal activations to data-driven, natural-language concepts, while an autoregressive decoder directly predicts model behavior. Our approach employs a two-stage paradigm—self-supervised pretraining followed by instruction fine-tuning—and introduces an automatic evaluation metric (auto-interp score) to optimize bottleneck quality. Results: Experiments demonstrate significant improvements over baselines across diverse tasks—including jailbreak detection, implicit prompt identification, latent concept injection, and user attribute inference. The learned concept representations exhibit strong cross-task generalization, and both bottleneck quality and downstream performance scale consistently with data volume.

Existing concept-based interpretability methods for CNNs overlook the semantic role of individual filters and the dynamic propagation of concepts across layers. To address this, we propose ConceptFlow—a novel framework that introduces a concept transition matrix to explicitly model inter-layer concept evolution, coupled with a filter-level concept attention mechanism to hierarchically and structurally dissect the generation and propagation pathways of high-level semantic concepts. ConceptFlow integrates concept activation analysis, attention-guided reasoning, and concept path tracing to explicitly reveal the hierarchical inference logic within CNNs. Experimental results demonstrate that ConceptFlow yields explanations with strong semantic consistency and high fidelity, significantly enhancing both the depth of understanding and the granularity of interpretability for CNN decision-making processes across multiple benchmark models and datasets.

This work addresses the lack of causal interpretability in black-box AI models by proposing a causal concept-driven explainable AI framework. The method employs post-hoc semantic concept extraction, constructs a causal graph between concepts and model outputs, and quantifies the effect of concept interventions on predictions via probabilistic sufficiency analysis—yielding faithful explanations with both local and global perspectives. Its key innovation lies in explicitly modeling the causal effects of concept interventions, ensuring explanations are both human-intelligible (high comprehensibility) and strictly consistent with the original model’s behavior (low fidelity loss). Experiments on CelebA demonstrate that the generated concept-based explanations exhibit clear semantics, strong readability, and classification performance highly aligned with the original model, empirically validating the framework’s effective balance between fidelity and interpretability.

This work addresses the challenge of efficiently embedding interpretable and controllable internal representations within large language models without resorting to complex feature identification and intervention procedures. The authors propose introducing a lightweight auxiliary loss during fine-tuning to designate 16 fixed dimensions in the residual stream as inert yet semantically meaningful concept indicators. By leveraging the model’s inherent pressure for computational efficiency, this approach encourages the model to actively rely on these pre-defined markers during generation, thereby automatically suppressing redundant encoding. Experiments demonstrate that this method enables precise and controllable generation interventions within a 3072-dimensional residual stream, validating the feasibility of constructing interpretable and manipulable internal representations through efficiency-driven training dynamics.