acceptodds
Under review as a conference paper at ICLR 2027

MOSAIC: Mode-Aware Prototypical NeuroSymbolic AI for Concept Learning

Abstract

NeuroSymbolic (NeSy) predictors combine the expressive power of neural networks with the safety and interpretability of symbolic methods by reasoning over learned intermediate concepts. Yet state-of-the-art methods remain vulnerable to reasoning shortcuts (RSs): incorrect input-to-concept mappings that nevertheless achieve high task accuracy under distant supervision. Many mitigation strategies require dense concept annotations, while recent prototypical approaches avoid both shortcuts and dense supervision by grounding each concept in a single prototype. However, they implicitly assume that each concept induces a single-mode distribution in the embedding space. In this paper, we show that the unimodality assumption gives rise to previously overlooked RSs at the latent-mode level, which we term implicit reasoning shortcuts (IRSs), in contrast to explicit reasoning shortcuts (ERSs) operating at the concept level. We prove that mode-blind predictors are structurally susceptible to IRSs: their optima hurts concept accuracy by conflating distinct modes and this ambiguity persists even with full concept supervision and complete data coverage. To address this limitation, we introduce ode-Aware Prtotypical Neuroymbolic for oncept Learning (), which discovers latent concept modes from unlabelled support data using infinite mixtures. We derive its embedding updates, characterise its deterministic optima, and establish, through an identifiability analysis, that optimal mode-aware predictors recover the intended concept semantics up to a permutation of latent modes. Empirically, MOSAIC substantially improves concept accuracy over existing reasoning-shortcut mitigation methods on the new Omniglot-EvenOdd benchmark and on the real-world, safety-critical BDDOIA task.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
NeuroSymbolic AI, Reasoning Shortcuts, Prototypical Networks, Identifiability theory

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Q2RkOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces MOSAIC, a neurosymbolic model that represents each concept by multiple prototypes instead of a single one. The model defines "Implicit Reasoning Shortcuts" as predicting correct concepts and task labels but "mispredicting" the intended latent mode, and provides theory for recovering these modes.

Critical strengths

The paper presents an interesting combination of neurosymbolic AI and Infinite Mixture Prototypes, which allows the model to adaptively discover multiple prototypes for each concept.

Critical weaknesses

The central motivation of the paper is that each concept may require multiple modes in latent space and that existing prototypical NeSy approaches do not support this.

The two main critical weaknesses are that (1) the importance of recovering (and existence of) these modes is unclear and the theoretical motivation is not convincing, and (2) the characterization of prior work is misleading, since many existing models already use multiple prototypes per predicted variable or concept.

Regarding (1):

  • The distinction between modes and ordinary variation around modes is unclear. Already in the running examples, it is unclear why different representations of writing a number (e.g. "3" vs "III") should be different modes, while different handwriting styles (e.g. "7" with or without a bar) should be variations near a single mode. Similarly, why pedestrians and bicycles could simply not be different variations of one mode for the concept "obstacle". This choice seems arbitrary. There is also no general expressivity argument, as a sufficiently expressive encoder can map all such variations near a single mode. If the different "desired modes" have independently justified human semantics, then the authors should explain why they should not instead be represented as distinct concepts (e.g. "bicycle", "pedestrian", "roman 3", "normal 3").
  • The problem seems unrelated to reasoning shortcuts. The proposed definition of an Implicit Reasoning Shortcut (IRS) does not really define a reasoning shortcut. An IRS can occur even when the model predicts both the explicit concepts and the downstream task perfectly, even when there are no logic rules in the architecture, etc. Reasoning shortcuts are a form of misalignment that can happen because the supervision signal of the task combined with the logic may be insufficient to learn the right concepts.
  • Additionally, Definition 2.1 is extremely brittle because it relies on exact equality of conditional distributions, which is very unlikely to hold in practice. There are many reasons why these distributions may differ that have nothing to do with reasoning shortcuts. For instance, any miscalibration of the model or approximation error would be classified as a "reasoning shortcut" according to this definition.

Regarding (2):

  • The paper’s statements about prior prototype work is incomplete and misleading. PNet (Andolfi and Giunchiglia, 2025) indeed uses a single prototype per explicit concept, but many existing works do not have this limitation. In particular, the authors seem to claim that PGCM (Colamonaco et al., 2026) has the single-mode problem, while this is factually wrong, as PGCM seems to learn multiple prototypes per concept. PGCM even seems to define each concept explicitly as a disjunction over multiple prototypes. More generally, many prototype-based networks use multiple prototypes to predict a single variable, e.g. ProtoPNet (Chen et al., 2019). Therefore, the paper seems to primarily address a limitation from the PNet architecture rather than a general limitation of the literature, which significantly narrows the scope and novelty of the contribution.
Questions for authors

I refer to the weaknesses mentioned above.

15 comments
AuthorsOpenReview ↗

Comment (1) — Modes relevance and their distinction from stylistic variations.

We thank the reviewer for their time and address the points raised below.

The content we present under each point will be included in the next version of the paper to improve its presentation.

We first answer the reviewer’s comment (1) in four points below.

Regarding comment (1)

1. On the importance of recovering the modes.

Learning latent fine-grained structure from coarse-grained supervision is a well-established problem in AI. Many learning problems involve an underlying hierarchy of concepts in which annotations are available only for coarse-grained categories, while the finer-grained categories remain unlabelled and must be inferred from data. This problem has been studied across several areas of AI. In inductive logic programming, target concepts can be represented as disjunctions of more specific sub-concepts [3], which can themselves be invented from examples of higher-level predicates [2]. Similarly, in computer vision, categories are often organised into coarse-to-fine hierarchies [4], while unannotated subpopulations within a class give rise to the well-studied problem of hidden stratification [5]. In natural language processing, learning fine-grained categories from coarse-grained supervision has also been explored as a means of reducing annotation costs [6]. MOSAIC addresses this problem in the neurosymbolic setting, where supervision is provided through symbolic knowledge defined over explicit (coarse-grained) concepts, while their underlying finer-grained realisations remain implicit. Rather than requiring these finer-grained categories to be separately annotated and introduced into the symbolic vocabulary, MOSAIC aims to automatically discover and represent them as latent modes with no additional supervision, while preserving the original explicit concepts used for reasoning.

2. On the distinction between modes and ordinary variations.

The distinction between modes and stylistic variations is formalised by our generative assumptions. Following [1], our mode-aware generative process (Figure 2) assumes that inputs are generated as x=ϕτ(m,s)\mathbf{x}=\phi_\tau(\mathbf{m},\mathbf{s}), where ϕτ\phi_\tau is invertible and smooth with respect to the style variables s\mathbf{s} (Assumption [A2]). Smoothness implies that continuous changes in style produce continuous changes in the input. Consequently, for a fixed mode m\mathbf{m}, whenever the support of S\mathbf{S} is connected, the set of generated inputs Xm=ϕτ(m,s)∣s∈supp(S)\mathcal{X_m} = \\{ \phi_\tau(\mathbf{m},\mathbf{s})\mid \mathbf{s}\in\mathrm{supp}(\mathbf{S}) \\} is also connected. Moreover, invertibility ensures that distinct modes generate disjoint sets of inputs, i.e., Xm∩Xm′=∅\mathcal{X_m} \cap \mathcal{X_{m^{'}}}=\emptyset for m≠m′\mathbf{m} \neq \mathbf{m'}.

This provides a principled interpretation of the distinction: stylistic variations describe continuous changes within a fine-grained category, whereas modes represent discrete factors of variation that cannot be explained by changes in style alone. For instance, variations in the stroke width or slant of an Omniglot\texttt{Omniglot} character can be captured by s\mathbf{s}, whereas Arabic 3 and Roman III represent distinct fine-grained categories associated with the same explicit concept. Similarly, buses and construction vehicles constitute different possible realisations of the explicit concept obstacle in BDD-OIA\texttt{BDD-OIA}. Observe these are example meant to provide an intuition to the reader. However, in the method we do not uniquely determine which fine-grained distinctions should be modelled as modes: this depends on the underlying generative factors and the chosen granularity of the concept hierarchy. Given such a decomposition, MOSAIC provides an inductive bias for preserving its heterogeneous structure through multiple prototypes, rather than requiring all realisations of an explicit concept to be represented by a single prototype. We will clarify this distinction and its underlying assumptions in the revised manuscript.

AuthorsOpenReview ↗

Comment (1) — Expressivity argument and definition of Implicit Reasoning Shortcuts.

3. On the importance of having such distinction in presence of a powerful encoder.

The reviewer claims:

“There is also no general expressivity argument, as a sufficiently expressive encoder can map all such variations near a single mode.”

We need to clarify that this is not true: this point has been addressed in Section 5.3. There we report a research question explicitly dedicated to the expressivity argument the reviewer refers to. The results already show that a deeper encoder can be beneficial for improving the concept F1-Macro score in Omniglot-EvenOdd\texttt{Omniglot-EvenOdd} (while remaining below MOSAIC’s), but it immediately backfires on BDDOIA\texttt{BDDOIA}. Indeed, a deeper encoder quickly overfits the limited support available and exhibits poor generalisation over the complex street scenes featured in the task. We further elaborate on this point below, and we provide additional results obtained with a deeper encoder (ResNet18) than the one reported already in the submitted manuscript (ResNet12). The settings are the same as Table 22 (BDD-OIA\texttt{BDD-OIA}, DPL baseline, support-set size 20 per concept).

DPLAcc(C) ↑F1(C) ↑Acc(Y) ↑F1(Y) ↑Cls(C) ↓
+PNet72.52 ± 1.6154.81 ± 0.9358.41 ± 2.6143.80 ± 1.790.00 ± 0.00
+ResNet12-PNet50.16 ± 5.9839.52 ± 4.9547.40 ± 0.1325.88 ± 0.250.07 ± 0.08
+ResNet18-PNet45.87 ± 8.2836.01 ± 6.1547.38 ± 0.0925.80 ± 0.090.13 ± 0.08
+MOSAIC89.13 ± 0.1257.59 ± 1.2962.45 ± 3.0241.42 ± 3.170.00 ± 0.00

As we can see, the more expressive encoder further degrades the performance of the PNet with the standard four-stack convolutional encoder and the shallower ResNet12 encoder. In particular, the mean concept collapse Cls(C) we obtained with the ResNet18 nearly doubles the one of ResNet12. Notice that from a theoretical perspective, by Corollary 4.7, no increase in capacity prevents an encoder that maps all modes onto the same prototype from falling into an IRS.

4. On the definition of Implicit Reasoning Shortcuts.

Our Definition 2.1 directly generalises the traditional definition of reasoning shortcuts [1] to the level of latent modes. Specifically, we distinguish between explicit reasoning shortcuts (ERSs), corresponding to traditional RSs where incorrect explicit-concept predictions nevertheless maximise the task likelihood, and implicit reasoning shortcuts (IRSs), where the predictor fails to recover the intended latent modes, up to permissible permutations. When each explicit concept has a single mode, our definition reduces to the traditional one. More generally, every ERS implies an IRS, but the converse does not hold.

The reviewer's observation that "an IRS can occur even when the model predicts both the explicit concepts and the downstream task perfectly" is correct and, indeed, precisely the additional phenomenon our generalisation is intended to capture.

The connection to traditional RSs lies in the semantic ambiguity left unresolved by the available supervision. In conventional RSs, the task supervision combined with the symbolic knowledge is insufficient to uniquely determine the intended explicit concepts. In our setting, this ambiguity extends to the underlying modes: the symbolic knowledge is defined over explicit concepts, and the downstream likelihood depends on the modes only through the taxonomy τ\tau mapping them to explicit concepts. Consequently, mode assignments preserving the explicit concepts are indistinguishable to the downstream supervision. Thus, even when the task supervision is sufficient to recover the correct explicit concepts, it need not identify their underlying modes. Proposition 4.1 characterises the resulting deterministic optima, while Corollary 4.7 establishes that mode-blind prototypical predictors cannot recover the intended latent-mode structure, even under complete mode coverage.

Importantly, the relevance of IRSs goes beyond recovering latent structure. Although an IRS does not necessarily imply incorrect explicit-concept predictions, the mode-blindness underlying these failures can also compromise explicit-concept learning. By representing heterogeneous realisations through a single prototype, mode-blind predictors impose a restrictive inductive bias that can lead to incorrect explicit-concept assignments under limited supervision. Our existing experiments already demonstrate the practical relevance of this distinction. For example, on Omniglot-EvenOdd (10 alphabets), Table 3 shows that CCN+ with MOSAIC improves mode purity from 2.93% to 80.98% and reduces IRS-rate from 97.14% to 2.94%, while Table 1 shows a simultaneous improvement in explicit-concept F1 from 65.12% to 87.12%. Furthermore, Figure 5 shows that increasing support coverage allows MOSAIC to discover additional modes, accompanied by improvements in explicit-concept accuracy.

AuthorsOpenReview ↗

Comment (1) — Additional experimental analysis about the Implicit Reasoning Shortcuts definition.

The results we included in the answer above demonstrate that modelling the finer-grained structure is not merely an additional representational objective, but provides an effective inductive bias for improving the performance, as measured by accuracy and F1 score on the explicit concepts on which symbolic reasoning operates.

Additional empirical analysis.

To further illustrate why our generalisation is meaningful, we conducted additional experiments on Omniglot-EvenOdd\texttt{Omniglot-EvenOdd} measuring how frequently IRSs coexist with correct explicit-concept and downstream task predictions. Specifically, we introduce two metrics:

a. Concept accuracy under IRS, (Acc(C∣IRS))Acc(\mathbf{C}\mid\mathrm{IRS})): the proportion of IRS-affected images for which the explicit concept is correctly predicted.

b. Joint concept–task accuracy under IRS, (Acc(C,Y∣IRS))Acc(\mathbf{C},\mathbf{Y}\mid\mathrm{IRS})): the proportion of IRS-affected images for which both the explicit concept and downstream label are correctly predicted.

ModelMode purity (global) ↑IRS-rate ↓Query imgs in an IRS ↓Acc(C | IRS) ↑Acc(C,Y | IRS) ↑
NN1.61 ± 0.3198.57 ± 0.4999.85 ± 0.1927.29 ± 5.3623.51 ± 4.66
SL1.13 ± 0.0799.14 ± 0.0099.96 ± 0.138.47 ± 2.565.60 ± 2.34
Scallop +C+E+R+PRE0.64 ± 0.3599.60 ± 0.26100.00 ± 0.0015.79 ± 3.4010.90 ± 1.55
CTSketch +C+E+R+PRE0.58 ± 0.1999.63 ± 0.13100.00 ± 0.0016.07 ± 3.9810.61 ± 1.90
CCN+ +C+E+R+PRE1.10 ± 0.2499.23 ± 0.1899.96 ± 0.1328.76 ± 8.028.46 ± 6.41
Scallop + PNet3.00 ± 0.0697.14 ± 0.00100.00 ± 0.0069.90 ± 1.1448.25 ± 1.72
CTSketch + PNet3.03 ± 0.0497.14 ± 0.00100.00 ± 0.0070.28 ± 1.0848.63 ± 1.49
CCN+ + PNet3.02 ± 0.0597.14 ± 0.00100.00 ± 0.0069.21 ± 0.8947.42 ± 1.27
Scallop + MOSAIC77.56 ± 1.253.66 ± 1.377.09 ± 2.8590.32 ± 3.1179.56 ± 2.93
CTSketch + MOSAIC77.72 ± 0.993.66 ± 0.746.92 ± 1.4690.92 ± 4.1280.20 ± 3.51
CCN+ + MOSAIC77.69 ± 1.272.94 ± 0.515.70 ± 1.3091.49 ± 3.3981.18 ± 2.49

All values are percentages (mean ± standard deviation over 10 runs). These results substantiate the reviewer's observation: for CTSketch + PNet, 70.28% of IRS-affected images have the correct explicit concept and 48.63% have both the correct concept and downstream label. Such latent-mode errors remain invisible to conventional concept and task accuracy metrics, despite being captured by Definition 2.1. Moreover, mode-blind predictors map 350 underlying characters into at most 10 predicted categories, yielding an IRS-rate of at least (1-10/350=97.14%), consistent with the limitations characterised by Corollary 4.7. Conversely, MOSAIC substantially reduces this rate by explicitly representing multiple modes per concept.

We also agree that latent-mode misalignment can arise without a symbolic reasoning layer, as illustrated by the additional NN baseline, realising a knowledge-agnostic version of the Semantic Loss baseline (same architecture, cross-entropy loss over the downstream label rather than the semantic objective). Our claim is not that logical reasoning necessarily causes IRSs, but that the semantic ambiguity studied in traditional RSs extends to the latent structure underlying explicit concepts. Our theoretical analysis characterises this ambiguity, while our experiments demonstrate that explicitly modelling such structure can also improve the recovery of the explicit concepts required for reasoning.

We will revise the manuscript to clarify this connection and the distinction between ERSs and IRSs.

While we believe the terms implicit and explicit reasoning shortcut captures the generalisation of traditional RSs, we would welcome alternative terminology if the reviewer feels it would make both this distinction and their connection clearer.

Reviewer_SZoHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper explores the issue of "reasoning shortcuts" in Neuro-Symbolic AI, which arise when a concept possesses multiple underlying patterns. The authors distinguish between Explicit Reasoning Shortcuts (ERS) and Implicit Reasoning Shortcuts (IRS) and propose the MOSAIC method, which employs multiple prototypes to represent each concept. Furthermore, the authors provide a theoretical analysis of concept identifiability and demonstrate the performance improvements achieved by this method on the Omniglot-EvenOdd and BDD-OIA datasets.

Critical strengths
  • The paper is well-motivated. The single-prototype assumption of existing prototypical NeSy models is indeed somewhat restrictive.
  • The distinction between ERSs and IRSs is interesting, and the theoretical analysis provides insights into concept identifiability.
  • MOSAIC has demonstrated promising results across different NeSy frameworks.
Critical weaknesses
  • The proposed method primarily involves Infinite Mixture Prototypes (Allen et al., 2019) into existing prototypical-based NeSy models. The authors should further clarify the novelty of this approach in the literature of NeSY (maybe it not that new in the original literature).
  • The identifiability results rely on strong assumptions, specifically mode clusterability and complete mode coverage.
  • The experimental evaluation of IRS mitigation strategies is limited. On the BDD-OIA dataset, the improvement in concept F1 scores is marginal, and there is no direct evaluation of the IRS phenomenon itself. Furthermore, the practice of selecting two labeled samples for each character in the Omniglot dataset appears to presuppose prior knowledge of the underlying modes.
Questions for authors
  1. Could the authors provide an example illustrating how IRSs are harmful when explicit concept predictions are correct?
  2. In the absence of knowledge regarding the underlying modes, how is complete coverage of all modes achieved?
10 comments
AuthorsOpenReview ↗

On the novelty of MOSAIC in the neurosymbolic literature.

We thank the reviewer for their positive assessment and for their constructive comments, which allow us to better position our contribution, clarify the role of our assumptions and strengthen the empirical evaluation around implicit reasoning shortcuts. We answer each weakness and question below, and we will incorporate these clarifications and the new results in the revised manuscript.

Regarding Weakness (1)

MOSAIC builds upon Infinite Mixture Prototypes (IMP [1]) to discover modes automatically from the data, leveraging a Chinese Restaurant Process approach. What is new is not the mixture itself, but its integration in the neurosymbolic setting and the study arising from it.

We detail each aspect below.

1. A failure mode that NeSy predictors could not express so far.

The NeSy literature on reasoning shortcuts ([2,3,4]) has so far not considered the consequences of learning latent fine-grained structure from coarse-grained labels under the distant supervision of symbolic knowledge prone to reasoning shortcuts. In Section 2 we fill this gap by generalising the data generation process in [2] to include multiple implicit concepts — the unlabelled fine-grained latents — and Definition 2.1 accordingly separates two different kinds of reasoning shortcuts. IMP, being a few-shot classifier trained without any background knowledge, operates in a completely different setting.

2. Coupling mode discovery with the symbolic supervision.

IMP classifies each query through the softmax over its distances to the closest cluster of each class, and it is trained on episode labels only. In MOSAIC, instead, (i) the distribution over the explicit concepts used for symbolic reasoning marginalises over every mode of the mode-lift τi−1(c)\tau_i^{-1}(c) (Eq. 4–5), so that any NeSy layer (DPL, SL, Scallop, CTSketch, CCN+) can consume it unchanged, and (ii) the NeSy loss is computed on this marginal, so that the knowledge reaches every mode in proportion to its responsibility. Theorem 3.1 characterises how the resulting gradient decomposes into mode-specific updates around a mode-aware belief centre, which has no counterpart in IMP [1].

3. Guarantees specific to the NeSy setting.

Proposition 4.1 counts the deterministic optima of mode-aware predictors under the knowledge K\mathsf{K}, generalising the shortcut count of [2,4]. The connection with identifiability theory established in Theorem 4.4, Lemma 4.5 and Theorem 4.6 lifts the guarantees on intended concept learning in the NeSy setting to families of mode-aware predictors that are equivalent up to Euclidean transformations. All these results concern the interplay between prototypes and symbolic knowledge, which falls outside the scope of IMP and previous literature.

4. A benchmark for implicit reasoning shortcuts.

Omniglot-EvenOdd\texttt{Omniglot-EvenOdd} is, to the best of our knowledge, the first NeSy learning and reasoning benchmark in which each explicit concept is realised (in both its 10- and 40-alphabet variants) by tens of visually disjoint modes, which makes IRSs directly measurable. We will make this positioning explicit in the Introduction through a dedicated list of contributions.

AuthorsOpenReview ↗

On the assumptions behind the identifiability results.

We agree with the reviewer that mode clusterability [A3] and complete mode coverage (Definition 4.2) are strong assumptions, but they are needed only for the strongest form of our theoretical guarantees, i.e., learning concepts with the intended semantics up to mode permutation (Theorem 4.6). We clarify their role below.

1. On mode clusterability.

[A3] transposes the clusterability assumption of [4] from the explicit concepts to their latent modes. Unlike in [4], however, the induced nearest-prototype map γi ⁣:[mi]→[mi]\gamma_i \colon [m_i] \to [m_i] is not required to be injective, i.e. distinct modes may share the same prototype. Thus, this assumption is not stronger than the corresponding one in previous literature. Importantly, the extent to which [A3] holds can also be measured empirically, as we do through mode purity (Table 3 and Tables 24–25 of the paper), which allows us to verify how well the model adheres to [A3].

2. On complete mode coverage.

This assumption is required only for Theorem 4.6, i.e. for the guarantee of recovery of the intended concept semantics. The remaining results do not rely on it: Theorem 3.1 (mode-specific embedding updates), Proposition 4.1 (deterministic optima, which depend on the support only through the set SisupS_i^{\mathrm{sup}} of annotated modes), Theorem 4.4 and Lemma 4.5 (identifiability of the prototype extractors). Notably, the negative result of Corollary 4.7 holds whatever the mode coverage of the support set: no amount of support data allows a mode-blind predictor to recover the modes. Also, complete coverage is not only sufficient but also necessary for mode-awareness in prototypical predictors: prototypes are created only from support datapoints, and no prototype can be expected to represent a mode that has never been observed. We would also like to remark that the paper devotes a dedicated research question to the regime without complete coverage (RQ4, Section 5.4 and Appendix D.4). On Omniglot-EvenOdd\texttt{Omniglot-EvenOdd}, as coverage decreases MOSAIC discovers fewer modes and degrades gracefully onto the mode-blind PNet. On BDD-OIA\texttt{BDD-OIA}, where the modes are unknown and coverage cannot be verified, MOSAIC still improves monotonically with the support size and at S=5S=5 already shows improvements over PNet in both concept and task accuracy (cf. Table 19 of the paper). Hence, while complete coverage is needed for the identifiability guarantee, the empirical benefits of MOSAIC are clear even when it holds only partially.

AuthorsOpenReview ↗

Direct evaluation of Implicit Reasoning Shortcuts across all mitigation strategies.

Regarding Weakness (3)

We thank the reviewer for this suggestion, which led us to considerably extend our analysis of IRSs. We address separately: (A) the direct evaluation of the IRS phenomenon across all the mitigation strategies and (B) the improvement in concept F1 on BDD-OIA\texttt{BDD-OIA}.

A. Direct evaluation of the IRS phenomenon.

The submitted paper already measures IRSs directly through mode purity and IRS-rate =1−∣im(γ^i)∣/mi= 1 - |\mathrm{im}(\hat{\gamma}_i)|/m_i, i.e. the fraction of true modes whose empirical image collides onto the same prototype (Table 3 and Tables 24–25 of the paper). We agree, however, that these diagnostics were reported only for MOSAIC and PNet. We have therefore extended them to every mitigation strategy reported in the paper, i.e. entropy (E) regularisation, concept supervision (C), pretraining (PRE), reconstruction (R) and all their combinations, for three reasoning layers (CCN+, CTSketch and Scallop), together with PNet and ResNet-PNet. Alongside purity and IRS-rate, we report the fraction of query images falling in an IRS, together with two metrics quantifying how often an IRS goes unnoticed by the standard concept and task metrics:

a. Concept accuracy under IRS, Acc(C∣IRS)Acc(\mathbf{C}\mid\mathrm{IRS}): the concept accuracy score restricted to the datapoints whose mode representation collides with another one.

b. Joint concept–task accuracy under IRS, Acc(C,Y∣IRS)Acc(\mathbf{C},\mathbf{Y}\mid\mathrm{IRS}): among the same datapoints in a., the fraction of them receiving both the correct explicit concept and the correct sum.

For the non-prototypical baselines, each concept is mapped to its own digit by the logic layer and plays the role of a prototype in γ^i\hat{\gamma}_i. The results are reported in Table 1 below (CCN+) and in Tables 2–3 in the next comment (CTSketch and Scallop).

ModelMode purity (global) ↑IRS-rate ↓Query imgs in an IRS ↓Acc(C | IRS) ↑Acc(C,Y | IRS) ↑
CCN+0.84 ± 0.1799.37 ± 0.1199.98 ± 0.0613.25 ± 6.007.05 ± 3.94
+C0.92 ± 0.2199.37 ± 0.1799.94 ± 0.1320.87 ± 4.427.79 ± 2.29
+E1.21 ± 0.4599.06 ± 0.36100.00 ± 0.0015.08 ± 5.879.99 ± 4.67
+PRE0.76 ± 0.1799.43 ± 0.1399.92 ± 0.1014.35 ± 6.248.61 ± 3.58
+C+E1.48 ± 0.4598.86 ± 0.4099.81 ± 0.2724.19 ± 6.5011.18 ± 3.02
+C+PRE1.05 ± 0.2099.31 ± 0.14100.00 ± 0.0031.54 ± 8.9012.57 ± 8.83
+E+PRE0.95 ± 0.5799.31 ± 0.48100.00 ± 0.0013.70 ± 3.898.73 ± 2.71
+C+E+PRE1.06 ± 0.2699.29 ± 0.1999.96 ± 0.1228.54 ± 7.087.51 ± 6.72
+R0.83 ± 0.1999.37 ± 0.17100.00 ± 0.0012.53 ± 3.197.89 ± 2.34
+R+C0.82 ± 0.1499.46 ± 0.09100.00 ± 0.0017.61 ± 3.667.61 ± 2.98
+R+E1.31 ± 0.6098.83 ± 0.6299.95 ± 0.1315.23 ± 5.2710.59 ± 6.54
+R+PRE0.66 ± 0.0599.46 ± 0.09100.00 ± 0.009.06 ± 0.364.55 ± 1.72
+R+C+E1.54 ± 0.4498.74 ± 0.5199.82 ± 0.2629.11 ± 7.1013.32 ± 2.90
+R+C+PRE0.97 ± 0.1799.37 ± 0.11100.00 ± 0.0027.60 ± 6.528.29 ± 6.71
+R+E+PRE0.77 ± 0.5999.40 ± 0.6299.98 ± 0.0612.38 ± 3.388.33 ± 3.32
+R+C+E+PRE1.10 ± 0.2499.23 ± 0.1899.96 ± 0.1328.76 ± 8.028.46 ± 6.41
+PNet3.02 ± 0.0597.14 ± 0.00100.00 ± 0.0069.21 ± 0.8947.42 ± 1.27
+ResNet-PNet3.05 ± 0.0497.14 ± 0.00100.00 ± 0.0078.78 ± 0.8562.05 ± 1.39
+MOSAIC77.69 ± 1.272.94 ± 0.515.70 ± 1.3091.49 ± 3.3981.18 ± 2.49

Table 1: Omniglot-EvenOdd\texttt{Omniglot-EvenOdd} (10 alphabets), CCN+. Mode-level diagnostics of MOSAIC against all the mitigation strategies and the mode-blind prototypical baselines on the held-out test queries. All values are percentages (mean ± standard deviation over 10 runs).

Reviewer_higoOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This work identifies the implicit reasoning shortcuts issue for neurosymbolic systems, formalizes it, can proposed MOSAIC as a solution. MOSAIC learns the concept from the latent space dynamically, such that more than one prototypes can be discovered for each concept. The number of prototypes is dynamically determined from the data via an application of the infinite mixtures approach. The authors provides both theoretical analysis of the properties of MOSAIC and empirical performance. Evaluated on the benchmark introduced by this work, Omniglot-EvenOdd, as well as on BDD_OIA autonomous driving tasks, MOSAIC show an improved quality on the learned concepts.

Critical strengths

S1: This work identifies the limitation of single-prototype based approaches, and proposed an effective method to learn multiple latent modes in the same concept representation. The authors formalize explicit and implicit reasoning shortcuts and provides substantial theoretical analysis on embedding updates, deterministic optima, and other characteristics of the proposed solution.

S2: The evaluations show strong empirical improvements. Evaluated on concept accuracy, MOSAIC consistently outperforms the other mitigation methods and single-prototype baselines. On the benchmark introduced by the authors, the improvements are particularly substantial, while on the BDD-OIA (more real life) experiments, the gain is still significant. The authors also provide relevant ablations on encoder capacity, support-set size, and unimodal settings. These results are convincing and provides empirical support for MOSAIC as the proposed solution.

Critical weaknesses

W1: Although MOSAIC largely improves concept accuracy, it is not clear how well it performs when extended to the downstream tasks, when making decisions for the actual actions. The main results shown are mostly concerning concept correctness, where action prediction performance is found in the Appendix where there seems to be a decrease when compared with baseline approaches. Is there a tradeoff between concept accuracy and downstream action accuracy?

Questions for authors

Q1: see W1, can the authors discuss more into how MOSAIC affect the downstream tasks when it comes to making a decision in real life scenarios?

4 comments
AuthorsOpenReview ↗

On the discrepancy between concept and downstream performance.

We thank the reviewer for their positive assessment and for giving us the opportunity to discuss the discrepancy between concept and downstream performance.

To better understand this discrepancy and investigate whether the learned concepts support generalisation beyond the training distribution, we

(i) conduct and report an additional experiment evaluating compositional generalisation to previously unseen concept combinations,

(ii) explain that the macro-F1 downstream score is lower on Omniglot-EvenOdd\texttt{Omniglot-EvenOdd} due to the particular evaluation setting of our benchmarks, which deliberately restricts the distribution of downstream labels to encourage reasoning shortcuts — as shown in Table 13 of the paper, MOSAIC already improves downstream accuracy (Acc(Y)) over the CCN+ baseline, 80.22% vs. 63.78%, despite achieving a lower macro-F1 score — and finally,

(iii) give an insight in the reason behind the lower downstream performance on BDD-OIA.

On the compositional generalization on Omniglot-EvenOdd\texttt{Omniglot-EvenOdd}: we created a new evaluation split by re-pairing test images so that the pairs present have either form (even, odd) or (odd, even). Notice that in both the training set and the evaluation set used in the paper we only have (even, even) and (odd, odd) pairs. This experiment thus aims to test the ability of the model to generalize to situations where the underlying concepts appear in combinations not seen in the training set. This is a very realistic scenario. For example, in autonomous driving, a model may encounter familiar objects and road conditions in combinations that were not represented during training. Correctly identifying the underlying concepts allows the symbolic reasoning component to determine the appropriate action without requiring every possible combination to have been observed.

The results of this additional evaluation are reported in Table 1, and clearly show that a CCN+MOSAIC model — leveraging concepts with the intended semantics — easily generalises to unseen sums, while CCN+ alone does not predict correct downstream labels anymore, as it relies on spurious concept associations. We will include these new results in the paper.

CCN+ MOSAICCCN+ baseline
Acc(C) (%)89.52 ± 0.5012.40 ± 4.46
Acc(Y) (%)80.21 ± 0.910.00 ± 0.00
F1(Y) (%)85.22 ± 0.740.00 ± 0.00
Table 1: Omniglot-EvenOdd\texttt{Omniglot-EvenOdd} (10 alphabets): Evaluation on previously unseen mixed-parity concept combinations, obtained by re-pairing test images.

On the downstream F1(Y). We now explain why MOSAIC achieves substantially higher downstream accuracy than CCN+ on the original Omniglot-EvenOdd\texttt{Omniglot-EvenOdd} benchmark, while recording a lower macro-F1 label score. The discrepancy in macro-F1 can be explained by the evaluation metric and the deliberately restricted downstream label distribution. Specifically, F1(Y) is macro-averaged over the union of true and predicted labels. As shown in Table 2 (included in the answer below), although only three sums ({6, 10, 12}) occur in the test set, MOSAIC predicts 18 distinct sums, whereas CCN+ concentrates almost exclusively on the three observed sums, with only 0.02% of its predictions falling outside this set.

Consequently, MOSAIC's macro-F1 is averaged over approximately 18 classes, 15 of which necessarily contribute zero because they never occur in the ground truth. This places an upper bound of 3/18 = 16.67% on its macro-F1 under the observed prediction support, independently of its accuracy on the three actual classes. In contrast, CCN+'s macro-F1 is averaged over only 3 or 4 classes, depending on the seed, and therefore does not suffer from the same dilution.

This explains why MOSAIC achieves substantially higher downstream accuracy (80.22% vs. 63.78%) but lower macro-F1 (14.30% vs. 35.65%). Importantly, the discrepancy does not indicate that MOSAIC is worse at predicting downstream labels: rather, its occasional concept prediction errors produce a wider range of incorrect sums, each of which contributes an additional zero to the macro-average, regardless of how rarely it occurs. We will clarify this in the paper.

AuthorsOpenReview ↗

On the discrepancy between concept and downstream performance Cont'd.

SumTrue label (%)CCN+MOSAIC predicted (%)CCN+ baseline predicted (%)
00.000.00 ± 0.000.00 ± 0.00
10.000.24 ± 0.080.00 ± 0.00
20.000.46 ± 0.110.00 ± 0.00
30.001.55 ± 0.220.00 ± 0.00
40.000.96 ± 0.220.00 ± 0.00
50.001.85 ± 0.250.00 ± 0.00
626.1423.78 ± 0.4317.12 ± 8.90
70.001.65 ± 0.160.00 ± 0.00
80.001.67 ± 0.170.02 ± 0.03
90.001.90 ± 0.230.00 ± 0.00
1055.5345.44 ± 0.4671.63 ± 19.45
110.001.65 ± 0.200.00 ± 0.00
1218.3414.59 ± 0.2911.23 ± 14.58
130.001.48 ± 0.210.00 ± 0.00
140.000.87 ± 0.100.00 ± 0.00
150.001.01 ± 0.280.00 ± 0.00
160.000.37 ± 0.110.00 ± 0.00
170.000.34 ± 0.120.00 ± 0.00
180.000.20 ± 0.090.00 ± 0.00
Outside the test labels0.0016.20 ± 0.880.02 ± 0.03

Table 2: Omniglot-EvenOdd\texttt{Omniglot-EvenOdd} (10 alphabets), test split (19,284 pairs). Percentage of the test predicted sum, mean ± std over 10 seeds. Bold rows are the only sums present in the test set.

On BDD-OIA\texttt{BDD-OIA}. To understand why DPL achieves higher downstream accuracy despite substantially lower concept accuracy, we studied the distribution of the action combination, reported in Table 3 below and noticed that this distribution is highly unbalanced, with "Stop" (alone) and "Forward" (alone) accounting for 32.90% and 21.02% of the actions combinations appearing in the test set, respectively (approximately 54% combined).

Action combinationTrain (%)Val (%)Test (%)
Stop33.6033.6632.90
Forward22.1422.5121.02
Forward + Right11.1111.5911.26
Forward + Left + Right9.4010.229.19
Forward + Left9.227.719.84
Stop + Right4.344.194.57
Stop + Left3.533.393.74
Forward + Stop1.471.191.22
Stop + Left + Right1.360.971.77
Left1.191.371.12
Right1.011.061.12
Forward + Stop + Right0.480.260.79
Forward + Stop + Left0.390.620.55
All four0.320.400.46
No action0.270.530.33
Left + Right0.170.350.13

Table 3: BDDOIA\texttt{BDDOIA}: 16,082 train, 2,270 validation and 4,572 test frames. Each frame is labelled with a set of the four actions.

According to the rules, these actions are triggered by the prediction of two specific concepts "Traffic sign" and "Follow traffic" for which:

  1. DPL obtains high recall, 74.90% and 88.38%, but low precision, 12.21% and 17.53% respectively,
  2. MOSAIC obtains low recall , 10.46% and 4.97%, but high precision 76.25% and 58.18%.

The predictions on these two triggering concepts lead MOSAIC act conservatively and predict "No action" in 22.45% of the cases, while DPL only predicts "No action" for 2.25% of the cases, much closer to the ground truth at 0.33%.

These results suggest that DPL's higher downstream accuracy may partly rely on frequent, often incorrect concept activations that happen to produce common actions, whereas MOSAIC's more accurate concept predictions do not always activate the conditions needed to recover the annotated actions.

Overall, these results highlight that concept accuracy and downstream accuracy measure different aspects of model performance. On Omniglot-EvenOdd\texttt{Omniglot-EvenOdd}, we show that learning more accurate concepts enables compositional generalisation to previously unseen concept combinations, where models affected by reasoning shortcuts fail. On BDD-OIA\texttt{BDD-OIA}, we observe that higher downstream accuracy on the original test distribution can coexist with less accurate concept predictions, particularly when frequent actions can be triggered by spurious concept activations. Together, these findings reinforce the importance of evaluating concept correctness alongside downstream performance, as high task accuracy alone does not guarantee that a model has learned the intended concept semantics or will generalise beyond the observed distribution.

Area_Chair_pYf7OpenReview ↗

Engaging with the authors

Dear reviewer,

the authors have posted a reply to your review. Please have a look and interact with the authors as needed.

Thank you, Your AC

Dear authors,

Thank you for your detailed responses and additional results. I think most of my concerns are addressed and i will therefore remain a positive recommendation for accept.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the discussion and of the revised manuscript.

Dear Area Chair and Reviewers,

As the discussion period concludes, we wish to thank the reviewers for their time and their suggestions.

Reviewers found the work well motivated and the single-prototype assumption of existing prototypical NeSy models restrictive (SZoH), the distinction between explicit and implicit reasoning shortcuts interesting (SZoH, higo), the theoretical analysis insightful (SZoH, higo), and the empirical improvements strong and consistent across reasoning layers, with convincing ablations (higo, SZoH). Reviewer higo confirmed that most of their concerns are addressed and maintained a positive recommendation.

Below we summarise how we addressed each set of comments. All changes are already in the revised PDF and are colour-coded by reviewer: green for Reviewer higo, yellow for Reviewer SZoH and blue for Reviewer Q2Rk.

Reviewer higo (green): concept vs. downstream performance

  • F1(Y) on Omniglot-EvenOdd. We explain that macro-F1(Y) is bounded by the fraction of predicted sums that actually occur in the test set. This is why MOSAIC attains higher Acc(Y) than CCN+ (80.22% vs. 63.78%) but a lower F1(Y) (Appendix D.1).

  • Compositional generalisation. We tested our method on a new split of unseen mixed-parity concept combinations (Table 10). There, CCN+ with MOSAIC reaches 80.21% Acc(Y), while the unmitigated CCN+ drops to 0.00%.

  • BDD-OIA. We analyse the action imbalance of the two concepts that trigger the most frequent actions. This shows that the baselines' higher downstream scores partly rely on frequent, often incorrect concept activations (Appendix D.1).

Reviewer SZoH (yellow): novelty, assumptions and direct evaluation of IRSs

  • Novelty. The Introduction now has an explicit list of contributions that positions MOSAIC with respect to Infinite Mixture Prototypes.

  • Assumptions. Section 4 clarifies that complete mode coverage is needed only for Theorem 4.6, and that the extent to which mode clusterability holds can be measured empirically through mode purity.

  • Annotated samples in Omniglot-EvenOdd. Section 5 clarifies that selecting annotated samples per character is part of the evaluation protocol only. During training, annotations carry the explicit concept alone.

  • Direct evaluation of IRSs. We extend mode purity and IRS-rate, together with two new metrics, Acc(C | IRS) and Acc(C,Y | IRS), to all 16 mitigation settings × 3 reasoning layers, plus PNet and ResNet-PNet (Tables 28–30). Every mitigation strategy stays at an IRS-rate of 97.43% or higher, while MOSAIC lowers it to 2.94–3.66%. Table 31 extends the BDD-OIA prototype-usage analysis to DPL and reports the number of effective prototypes (Appendix D.1).

  • F1(C) on BDD-OIA. A fully supervised linear probe trained on 100% of the concept labels (Table 9) shows that MOSAIC, using 1.63% of the annotations, closes 79% (DPL) and 83% (CCN+) of the gap between the unmitigated models and this reference.

Reviewer Q2Rk (blue): motivation, connection to reasoning shortcuts and prior work

  • Modes vs. style. Section 2 now formalises this distinction through Assumption [A2]. Modes are anonymous latent variables of the data-generating process, recovered up to permutation as in identifiability theory, and need not be human-interpretable or known a priori.

  • ERSs vs. IRSs. The Introduction and Section 2 clarify how the two relate. ERSs are failures of symbol grounding; IRSs are failures of symbol representation, which persist under full concept supervision and, with limited supervision, also degrade grounding.

  • Encoder capacity. Section 5.3 (RQ3) already examined this in the submitted manuscript with a ResNet12 encoder, and we added ResNet18 results during the discussion; both show that larger encoders do not close the gap in our low-supervision regime, which the Abstract and Introduction now state explicitly.

  • Related work. We clarified the sentence about Colamonaco et al. (2026).

  • Evaluation protocol. We document that test-time supports follow the standard protocol of prototypical networks. We also report that results are unchanged when support scenes are excluded from scoring or drawn from the training split (Appendix C.2).

We hope the revised manuscript addresses the reviewers' concerns, and we remain available for any further clarification.

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.