Concept Sidechannel Models: Accuracy and Interpretability

• Concept Sidechannel Models are frameworks that combine human-understandable concept bottlenecks with an auxiliary sidechannel to capture extra task-relevant information.
• They address the trade-off between interpretability and accuracy by allowing controlled leakage to compensate for incomplete concept representations.
• Architectural implementations like CREAM use graph-based masking and SIS regularization to balance reliance between concept channels and sidechannels, enhancing model fidelity and intervenability.

Searching arXiv for recent and relevant papers on Concept Sidechannel Models and closely related concept-based leakage work. arXiv search query: "Concept Sidechannel Models concept bottleneck sidechannel leakage interpretability" Concept Sidechannel Models (CSMs) are concept-based models in which prediction is mediated by human-understandable concepts but is additionally supplemented by a sidechannel that carries extra task-relevant information outside the concept bottleneck. In the recent literature, CSMs are presented as a response to the accuracy limitations of Concept Bottleneck Models (CBNMs): the bottleneck preserves interpretability by constraining prediction to concepts, but it also restricts information flow; the sidechannel relaxes this restriction, often improving predictive performance while creating an explicit accuracy–interpretability trade-off (Debot et al., 7 Oct 2025). Subsequent work has refined this picture by arguing that leakage through the sidechannel is not uniformly harmful, especially when concept incompleteness is the norm, and by proposing both architectural and probabilistic tools for measuring, regularizing, and exploiting sidechannel information in a controlled manner (Zarlenga, 9 Jun 2026).

1. Formal definition and probabilistic structure

A unified probabilistic CSM meta-model factors the joint distribution as

$$p(y,c,z,x) = p(x)\cdot p(c|x)\cdot p(z|x)\cdot p(y|c,z),$$

where $X$ denotes the input, $C$ the interpretable concepts, $Z$ the sidechannel, and $Y$ the task variable (Debot et al., 7 Oct 2025). Task inference is correspondingly written as

$$p(y|x) = \sum_{c,z} p(c|x)\cdot p(z|x)\cdot p(y|c,z),$$

with sums replaced by integrals for continuous variables. In deterministic neural instantiations, these distributions are implemented via Dirac delta functions, so inference reduces to a single evaluation at predicted values.

Within this formulation, CBNMs are the special case in which the sidechannel is absent and $p(y|c,z)$ reduces to $p(y|c)$. The meta-model is explicitly described as subsuming existing CSM variants as special cases, including Concept Residual Models (CRM), where $Z$ is an embedding and $p(y|c,z)$ is unconstrained, and Concept Memory Reasoner (CMR), where $X$0 is a one-hot mask selecting rules and $X$1 implements logic rules (Debot et al., 7 Oct 2025).

This formalization is closely aligned with a broader description of concept-based models as deep neural networks that ground predictions on representations aligned with human-understandable concepts. In that setting, the concept representation is typically written as $X$2, and the central issue is that $X$3 may encode information not strictly attributable to intended concept semantics (Zarlenga, 9 Jun 2026). A plausible implication is that CSMs should be understood not as abandoning the concept bottleneck, but as making explicit the additional information pathway that many concept-based systems already acquire implicitly.

2. Leakage as the defining phenomenon

The literature on CSMs treats leakage as the presence of information in the learned concept representation or auxiliary pathway that exceeds what is licensed by the annotated concepts. Two main forms are formalized. Inter-concept leakage occurs when a learned concept representation $X$4 contains information about another concept $X$5 beyond what is available in the ground-truth labels: $X$6 Task leakage occurs when $X$7 contains label-relevant information about $X$8 that is not attributable to the concept itself: $X$9 These inequalities frame leakage as excess mutual information relative to the intended concept semantics (Zarlenga, 9 Jun 2026).

The same work identifies several sources of leakage. One source is the use of bypasses or residual connections in the label predictor, written as $C$0, where $C$1 is an unconstrained transformation of the input. This augments the bottleneck and provides a direct route for non-concept, task-relevant information to reach the label predictor. The examples explicitly given are Hybrid CBMs, which add unsupervised activations in the bottleneck, and residual or augmented CBMs, which directly pass incomplete information. Leakage is also associated with architectural and training choices such as soft or embedding-based representations, non-hard bottlenecks, and joint training schemes (Zarlenga, 9 Jun 2026).

In this sense, the sidechannel in a CSM is both an architectural object and an information-theoretic one. It is architectural because it is realized as a distinct pathway $C$2 or residual latent; it is information-theoretic because its effect is defined by whether it carries task-relevant or concept-irrelevant information beyond the supervised concepts. This suggests that the boundary between an ordinary concept model with leakage and an explicit CSM is often a matter of whether that auxiliary flow is modeled and controlled directly.

3. Accuracy, concept incompleteness, and benign leakage

The conventional narrative treated leakage in concept-based models as undesirable because it appeared to make models less interpretable. More recent work argues that this view is ill-posed, that the evidence linking leakage to reduced interpretability is often inconclusive, and that eliminating leakage can lead to impractical models under common real-world constraints (Zarlenga, 9 Jun 2026). The key premise is concept incompleteness: prior work often assumes a complete representation satisfying

$C$3

but this is described as rarely true in practice.

Under incomplete concept sets, prohibiting leakage can reduce task fidelity and can also make test-time interventions less effective. For this reason, the literature introduces benign leakage, a controlled form of leakage that retains task-relevant information not present in the known concepts while localizing it so that interventions on concepts remain meaningful (Zarlenga, 9 Jun 2026). Formally, a concept representation $C$4 exhibits benign leakage if each component decomposes as

$C$5

where $C$6 is concept-aligned and $C$7 is residual or leaky, and two conditions hold.

The first is sufficiency: $C$8 meaning that $C$9 captures what the concepts miss. The second is localization: $Z$0 meaning that task-relevant information for each concept remains localizable in $Z$1, so concept interventions are effective (Zarlenga, 9 Jun 2026).

This reframing changes the interpretation of the sidechannel. In the earlier CSM narrative, the sidechannel primarily expresses a trade-off: higher accuracy at the cost of lower representation interpretability. In the benign-leakage narrative, the sidechannel can also serve as a compensatory mechanism for incomplete concepts, provided that its contribution is localized rather than allowed to dominate the prediction process.

4. Architectural realizations

One concrete architectural family is CREAM, or Concept REAsoning Models. CREAM explicitly encodes concept–concept and concept–task relationships to enforce a desired model reasoning, while integrating a regularized side-channel to achieve competitive task performance and keep high concept importance (Kalampalikis et al., 5 Jun 2025). The architecture is described as comprising a representation splitter, a concept–concept block, and a concept–task block.

If the backbone produces features $Z$2, the representation splitter partitions them into $Z$3, interpreted as concept exogenous variables for $Z$4 concepts, and $Z$5, interpreted as side-channel latent variables for task prediction. Concept–concept relationships are encoded by an adjacency matrix $Z$6, with masked input

$Z$7

and concept prediction written as

$Z$8

where $Z$9 (Kalampalikis et al., 5 Jun 2025).

Concept–task relationships are encoded by an adjacency matrix $Y$0 for $Y$1 task classes. The concept–task block takes both the concepts and the sidechannel, with

$Y$2

and classwise prediction

$Y$3

where $Y$4 is the class-specific side-channel latent (Kalampalikis et al., 5 Jun 2025). The design intent is explicit: the sidechannel supplements rather than overrides concept-based reasoning, and the graph structure severs undesired information flows, including those needed to handle mutually exclusive concepts.

The same paper regularizes the sidechannel by applying dropout to the entire sidechannel during training with probability $Y$5, thereby forcing the model to rely solely on concepts for prediction whenever the sidechannel is removed. The training loss is

$Y$6

where $Y$7 balances concept prediction accuracy against task prediction accuracy (Kalampalikis et al., 5 Jun 2025).

A related but more general training proposal is to optimize directly for benign leakage. The sufficiency regularizer $Y$8 is defined by minimizing the expected task loss when all concepts are intervened, so that the label predictor sees ground-truth concepts. The paper states that, under well-specified conditions, minimizing this intervention-based loss is equivalent to maximizing $Y$9, thereby achieving sufficiency (Zarlenga, 9 Jun 2026). Localization is described as harder to optimize directly, but minimizing $$p(y|x) = \sum_{c,z} p(c|x)\cdot p(z|x)\cdot p(y|c,z),$$0 is argued to encourage localization in practice through the simplicity bias of SGD and DNN architectures.

5. Quantifying sidechannel reliance

A central methodological advance is the introduction of the Sidechannel Independence Score (SIS), a metric for quantifying a CSM’s reliance on its sidechannel by contrasting predictions with and without sidechannel information (Debot et al., 7 Oct 2025). In default mode, the model uses the ordinary, input-conditioned sidechannel: $$p(y|x) = \sum_{c,z} p(c|x)\cdot p(z|x)\cdot p(y|c,z),$$1 In bottleneck mode, the sidechannel is replaced by a prior independent of the input: $$p(y|x) = \sum_{c,z} p(c|x)\cdot p(z|x)\cdot p(y|c,z),$$2 with the prior typically approximated by averaging sidechannel outputs over the training set (Debot et al., 7 Oct 2025).

Let $$p(y|x) = \sum_{c,z} p(c|x)\cdot p(z|x)\cdot p(y|c,z),$$3 be the default-mode prediction and $$p(y|x) = \sum_{c,z} p(c|x)\cdot p(z|x)\cdot p(y|c,z),$$4 the bottleneck-mode prediction. SIS is then defined as

$$p(y|x) = \sum_{c,z} p(c|x)\cdot p(z|x)\cdot p(y|c,z),$$5

with empirical estimate

$$p(y|x) = \sum_{c,z} p(c|x)\cdot p(z|x)\cdot p(y|c,z),$$6

The interpretation is operational: $$p(y|x) = \sum_{c,z} p(c|x)\cdot p(z|x)\cdot p(y|c,z),$$7 means predictions never rely on the sidechannel, while $$p(y|x) = \sum_{c,z} p(c|x)\cdot p(z|x)\cdot p(y|c,z),$$8 means predictions always depend on it (Debot et al., 7 Oct 2025). Hoeffding’s inequality is used to quantify statistical confidence in empirical estimates on unseen data.

SIS is paired with SIS regularization, which penalizes the divergence between output distributions in default and bottleneck modes. The training objective is presented in maximum-likelihood form as

$$p(y|x) = \sum_{c,z} p(c|x)\cdot p(z|x)\cdot p(y|c,z),$$9

where increasing $p(y|c,z)$0 encourages the model to minimize sidechannel influence and thereby increase SIS (Debot et al., 7 Oct 2025).

The CREAM framework contributes an additional metric, Concept Channel Importance (CCI), derived from SAGE values: $p(y|c,z)$1 where $p(y|c,z)$2 is the global feature importance of the concept channel and $p(y|c,z)$3 that of the sidechannel. Values near $p(y|c,z)$4 indicate concept reliance, values near $p(y|c,z)$5 indicate equal reliance, and CCI can be tuned via sidechannel dropout (Kalampalikis et al., 5 Jun 2025).

The measurement problem is therefore now articulated at two levels. SIS measures whether predictions change when the sidechannel is ablated, which directly targets representation interpretability. CCI measures the relative importance of the concept and sidechannel channels, which targets reliance balance. Together, these tools convert a previously informal trade-off into an explicitly optimizable object.

6. Empirical findings, misconceptions, and terminological scope

The empirical picture reported across these papers is consistent. State-of-the-art CSMs trained solely for accuracy exhibit low representation interpretability, and SIS regularization substantially improves interpretability, intervenability, and the quality of learned interpretable task predictors (Debot et al., 7 Oct 2025). Without explicit regularization, models may rely heavily on the sidechannel even when it is unnecessary. With SIS regularization, SIS increases and predictions become more concept-derived, usually at a modest cost in accuracy, yielding a clear Pareto frontier between interpretability and performance (Debot et al., 7 Oct 2025).

The benign-leakage results sharpen the interpretation of these findings. When CMs with sidechannels, including Hybrid CBMs or CEMs, are trained with the sufficiency regularizer $p(y|c,z)$6, they attain both high accuracy and intervenability even when concepts are incomplete. Well-conditioned leaky models are reported to be no less interpretable than non-leaky models under standard metrics such as concept fidelity, intervenability, and concept-class weight structure. The reported trends hold across multiple datasets, concept incompleteness regimes, and noise levels (Zarlenga, 9 Jun 2026). This directly contradicts the misconception that any leakage necessarily destroys interpretability.

CREAM adds a more structured empirical claim: it mainly relies on concepts while achieving task performance on par with black-box models, and the embedded concept–concept and concept–task relationships ease model interventions and mitigate concept leakage (Kalampalikis et al., 5 Jun 2025). In that framework, concept leakage is defined as the case in which a CBM’s task accuracy on predicted concepts, used as input to a separate classifier, exceeds that of an optimal predictor using the true concepts. The paper attributes leakage reduction to graph structure and masking in both concept–concept and concept–task blocks (Kalampalikis et al., 5 Jun 2025).

The abbreviation CSM is not unique across the arXiv literature. In medical imaging, “Deep Structural Causal Shape Models” also use the acronym CSM for causally grounded 3D shape models that combine geometric deep learning with structural causal models and support interventional and counterfactual reasoning about 3D meshes (Rasal et al., 2022). That usage is conceptually unrelated to Concept Sidechannel Models, but the acronym collision is relevant for literature searches and citation hygiene.

A broader terminological resonance comes from classical side-channel security. There, side channels are modeled as auxiliary outputs that help an adversary infer a secret, and maximal leakage is proposed as an operationally interpretable leakage metric that can upper bound adversarial gain more tightly than mutual information or channel capacity (Wu et al., 2020). A plausible implication is that the phrase “sidechannel” in Concept Sidechannel Models is apt not merely as metaphor: in both settings, an auxiliary pathway carries consequential information outside the intended semantic interface. The SoK literature on automated side-channel-aware design further emphasizes that modeling choices, abstraction levels, and evaluation metrics strongly affect what leakage can be detected and controlled (Buhan et al., 2021). For CSMs, the analogous lesson is that interpretability depends not only on whether a concept bottleneck exists, but on how auxiliary information flows are specified, localized, measured, and regularized.