acceptodds
Under review as a conference paper at ICLR 2027

Discovering Sparse Concept Graphs as Interpretable Surrogates for Vision Models

Abstract

Recent methods for mechanistic interpretability of vision models uncover interpretable concepts and trace the circuits underlying predictions for individual samples or classes. However, predictions for different classes often depend on shared concepts, and this shared structure is central to understanding how the network operates as a whole. We address this gap by learning sparse rules that describe how combinations of concepts in one layer give rise to concepts in the next, and compose these rules across layers into a ConcepTome, a single directed concept graph that serves as a global, symbolic surrogate of the network. The ConcepTome can be examined as a whole or queried for individual samples, single classes, or groups of classes to reveal the computational structure they share. We evaluate our approach on CLIP ViT and CLIP ResNet and show that the recovered ConcepTomes remain sparse while faithfully preserving model computation. Inspection reveals how concepts are combined, reused, transformed, and specialized throughout the network. Interventions show that the recovered structure isolates a compact set of concepts on which the computations and predictions of the model depend. To make this structure accessible, we develop an interactive interface for exploring the ConcepTome, visualizing concepts and rules, and intervening on predictions.

Primary area:
interpretability and explainable AI
Keywords:
Mechanistic Interpretability, Concepts, SAE, Rule Mining, Vision Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_BxFyOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper concerns connecting concepts (SAEs and their activation) between layers of a model through co-occurrence. The ensuing "Conceptome" is argued to serve as surrogate model "explanation".

Critical strengths

Important problem. Well argued attempt at combining SAEs into more global metacognitive framework.

Critical weaknesses

Unclear whether the resulting explanations are understandable, comprehensive and actionable. How are users supposed to make sense of the complex wiring patterns shown in the appendix C?

Questions for authors

Please give concrete use cases and worked examples of usability

2 comments

We thank the reviewer for recognizing the importance of the problem and the merits of our approach. We appreciate the insightful feedback regarding the usability of the CONCEPTOME and address this concern through concrete use cases and worked examples, which we have added to Appendix E (p. 20; Figures 12–17, pp. 24–27) of the revised manuscript.

W1 & Q1: Understanding and using the CONCEPTOME

The global structure illustrated in Appendix C allows users to explore individual rules and trace how concepts interact and propagate across layers. An important advantage of our approach, however, is that the CONCEPTOME can be queried for an individual sample, a class, or an arbitrary collection of samples. Users can thus extract subgraphs relevant to their questions and localize their analysis to specific layers or paths within the network. We illustrate this through the following use cases.

Understanding concept reuse and class differentiation. In addition to the CONCEPTOME example in Figure 6 of the original manuscript, we present class-specific paths for tabby (Figure 12, p. 24) and tiger cat (Figure 13, p. 24), illustrating how concepts combine across layers to reach their respective predictions. We further show a shared path that branches into distinct rules at later layers, leading to the two classes (Figure 14, p. 25). These examples demonstrate how users can investigate which concept interactions are reused across classes and where their computational pathways diverge, all within a common global structure.

Debugging prediction errors. By querying a collection of tiger cat images misclassified as tabby (Figure 15, p. 25), we identify a shared path associated with these incorrect predictions. This allows users to move beyond individual mistakes and investigate recurring concept interactions across samples exhibiting the same failure mode.

We further query collections combining correctly and incorrectly classified samples: balloon images misclassified as parachute together with correctly classified parachutes (Figure 16, p. 26), and rock beauty images misclassified as tench together with correctly classified tenches (Figure 17, p. 27). In both cases, we identify shared paths associated with correct and incorrect predictions. For instance, the balloon-to-parachute path contains concepts associated with both balloons and parachutes, providing insight into shared visual representations that may contribute to this confusion. These examples also illustrate how users can localize their perspective to particular portions of the network.

From explanation to intervention. The recovered paths provide starting points for investigating model behavior beyond visual inspection. Users can identify concepts and rules shared across failure cases, formulate hypotheses about their role in model predictions, and test these hypotheses through targeted interventions, as demonstrated in our interventional experiments in Section 5.

Together, these examples illustrate the CONCEPTOME as a queryable representation of model computation, enabling users to explore how vision models organize and reuse concepts, investigate shared and divergent computational pathways, and gain insights into the model's behavior across both successful and unsuccessful predictions.

 


Reviewer_BxFyOpenReview ↗

Not entirely convinced about the usability

Thank you for the response and attempt to address usability, however I still struggle with the complexity of the output and I am not yet convinced. I still think it is a quite cool idea and maintain my positive score.

Reviewer_VUeaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CONCORD, a method for discovering cross-layer concept graphs as interpretable surrogates for vision models. The methodology starts by learning sparse autoencoders (SAEs) at intermediate layers and then learns sparse rules describing how combinations of concepts in one layer predict concepts in the next. These rules are composed into a graph called Conceptome. The approach is evaluated on ImageNet subsets and CLIP ViT and ResNet backbones. They compare the approach with RRM, a recent sparse circuit discovery method for vision transformers.

Critical strengths

The proposed approach is interesting and ambitious. Learning concept rules between SAE representations at different layers can potentially provide useful insights into how concepts interact and how they are used by the model to make predictions.

Critical weaknesses
  1. The faithfulness of the learned rules is not sufficiently demonstrated. The main contribution of the paper is learning rules that describe how concepts of one layer combine to produce the concepts in the next, however the paper does not report how accurately these rules predict such concepts (i.e., per-concept accuracy, recall, precision, F1). Furthermore, the computational fidelity experiment in Table 2 does not include a matched random concept selection baseline. That would be useful to isolate how conceptome selection contributes to the metrics.

  2. It remains unclear how much of the model computation is explained by conceptome. The paper's experiments preserve the SAE reconstruction residual when performing interventions, which is reasonable. However, such residual can also contain information that is important for the model predictions, so retaining it makes it unclear how much of the recovered model behaviour is attributable to conceptome and how much to the residual. In other words, as I understand it, the interpretable surrogate needs two components: the conceptome, which is interpretable, and the SAE residual, which is not. But we do not know how much work each is doing, and therefore how much of the model we are explaining, if anything. Regarding this, I suggest the following sanity check: compute prediction agreement using only conceptome (retaining selected concept activations and suppressing non-selected ones), without SAE reconstruction residual. Then compute it using only SAE reconstruction residual (suppressing all SAE concept activations), and finally using both together. This would help understand how much of the original model behaviour is being explained by the interpretable component of the surrogate (conceptome) and how much by the uninterpretable one (SAE reconstruction residual).

Questions for authors

Could you perform the sanity checks suggested in W1 and W2? The results would materially affect my assessment.

7 comments

We thank the reviewer for their thorough assessment and insightful comments, which highlight important aspects of our work. Below, we address these points and present additional experimental results.

 


W1: Rule faithfulness and computational fidelity

We thank the reviewer for suggesting these evaluations. While predictive performance is generally informative, we would like to highlight an important finding from the development of CONCORD: recovering the underlying rule structure and accurately predicting individual concept activations are two distinct objectives.

1. Predictive performance and structural recovery. To investigate this distinction, we evaluate CONCORD on synthetic data with known ground-truth structures: 10,000 examples, 5,000 concepts in each of the input and output spaces, 30 planted conjunction rules with head and tail sizes of 5–15, and five uniformly sampled rules per example. We introduce noise by randomly flipping bits across the input and output matrices.

We measure structural recovery using Soft F1, which assigns partial credit to recovered rules based on the Jaccard overlap of their heads and tails with planted rules. Soft precision and recall average the best-match similarities from discovered to planted rules and vice versa, and Soft F1 is their harmonic mean.

The results of the synthetic noise ablation are reported below (mean ± standard deviation, %).

Noise (%)Prediction PrecisionPrediction RecallPrediction F1Soft F1
083.1 ± 1.897.3 ± 2.489.7 ± 1.893.2 ± 1.7
139.5 ± 2.140.0 ± 1.839.7 ± 1.787.5 ± 3.0
229.0 ± 1.525.3 ± 1.227.0 ± 1.385.3 ± 5.7
319.0 ± 1.018.4 ± 1.118.7 ± 1.082.7 ± 5.6
413.8 ± 0.916.1 ± 1.014.8 ± 0.986.6 ± 5.0
511.6 ± 0.618.0 ± 1.014.1 ± 0.681.8 ± 9.7

At 1% noise, comparable to the density of the underlying data, predictive F1 drops from 89.7% to 39.7%, while Soft F1 remains at 87.5%. Across 1–5% noise, mean Soft F1 remains above 80% despite predictive F1 ranging from 14.1% to 39.7%. These results demonstrate that CONCORD can recover the underlying structure even when predicting individual observed activations becomes challenging.

Having distinguished structural recovery from the prediction of individual concept activations, we now evaluate the latter following the reviewer's suggestion. We report micro-averaged precision, recall, and F1 for predicting downstream concept activations across all layer pairs of ViT-B/32 on ImageMeow, consistent with the running example used throughout the manuscript.

Layer pairPrecision (%)Recall (%)F1 (%)
183.195.989.1
266.281.773.1
358.468.463.0
453.067.959.6
545.878.057.7
646.774.557.4
755.270.461.9
853.173.061.4
954.170.061.0
1055.669.761.8
1146.972.857.0

While predictive F1 is lower in deeper layers, our synthetic findings illustrate why predictive metrics alone are insufficient to assess structural recovery. This is particularly relevant to CONCORD in this setting, where it learns dependencies over pooled, binarized concept representations, abstracting away activation magnitudes, spatial information, and SAE residuals.

 


W1: Rule faithfulness and computational fidelity (continued)

2. Matched random concept selection. To complement these predictive metrics with an assessment of the functional relevance of the recovered structure, we follow the reviewer's second suggestion and extend the computational fidelity evaluation of Table 2 with two matched random controls:

  • All-concept random: Randomly select concepts from the full concept vocabulary, matching the number of concepts selected by CONCEPTOME.
  • Active-concept random: Randomly select from observed active concepts, matching the number of active concepts retained by CONCEPTOME.

Both controls use the same intervention procedure and residual handling as our method. The second control additionally accounts for the activation status of the selected concepts.

The results across both backbones and all three subsets are reported below (mean ± standard deviation, %).

ModelSubsetMethodAgreementFaithfulnessCompleteness
ViT-B/32ImagenetteCONCEPTOME79.180.2100.0
All-concept random0.1 ± 0.13.3 ± 0.395.8 ± 0.2
Active-concept random0.9 ± 0.711.2 ± 0.5100.0 ± 0.0
ImageMeowCONCEPTOME79.692.9100.0
All-concept random0.0 ± 0.00.0 ± 0.0100.0 ± 0.0
Active-concept random1.4 ± 0.55.0 ± 0.3100.0 ± 0.0
ImagewoofCONCEPTOME64.886.7100.0
All-concept random0.3 ± 0.20.0 ± 0.0100.0 ± 0.0
Active-concept random1.8 ± 0.76.8 ± 0.2100.0 ± 0.0
RN50ImagenetteCONCEPTOME89.199.599.8
All-concept random0.1 ± 0.11.1 ± 0.398.6 ± 0.3
Active-concept random10.3 ± 2.126.2 ± 0.7100.0 ± 0.0
ImageMeowCONCEPTOME93.0100.097.1
All-concept random0.0 ± 0.02.2 ± 0.291.8 ± 0.2
Active-concept random7.4 ± 0.529.4 ± 0.398.7 ± 0.1
ImagewoofCONCEPTOME88.599.2100.0
All-concept random0.0 ± 0.00.7 ± 0.295.7 ± 0.3
Active-concept random5.9 ± 0.339.7 ± 0.5100.0 ± 0.0

Across both backbones and all three subsets, CONCEPTOME preserves 64.8–93.0% prediction agreement and 80.2–100.0% faithfulness, compared with at most 10.3% agreement and 39.7% faithfulness for the matched random controls. Importantly, this holds even when random selection is restricted to observed active concepts. These results demonstrate that the recovered model behavior cannot be attributed to the number or activation status of the retained concepts alone, highlighting the importance of the structure identified by CONCORD.

We additionally evaluate disappearance using target-fixed random interventions. Unlike the computational fidelity experiment, we keep the downstream targets identified by CONCEPTOME fixed and randomize the upstream concept selection under matched intervention budgets. This allows us to compare the influence of the learned antecedents and randomly selected concepts on the same downstream targets.

ModelSubsetCONCEPTOMEAll-concept randomActive-concept random
ViT-B/32Imagenette50.5443.45 ± 0.0349.61 ± 0.02
ImageMeow50.4743.76 ± 0.0249.35 ± 0.01
Imagewoof41.3433.60 ± 0.0441.15 ± 0.02
RN50Imagenette27.9715.87 ± 0.0226.10 ± 0.03
ImageMeow28.6513.72 ± 0.0125.18 ± 0.01
Imagewoof29.5414.59 ± 0.0226.39 ± 0.03

CONCEPTOME achieves higher disappearance rates than both random controls across all evaluated settings, with differences of 6.71–14.95 percentage points over the all-concept control and 0.19–3.47 over the active-concept control. The active-concept control is particularly challenging, as the randomly selected concepts may substantially overlap with the antecedents selected by CONCEPTOME. Nevertheless, CONCEPTOME consistently achieves higher disappearance rates across all evaluated settings.

Taken together, the matched random comparisons demonstrate that the high agreement and faithfulness achieved by CONCEPTOME are not simply a consequence of the number of concepts retained, but rather of the specific concepts and dependencies identified by CONCORD. These results support our objective of recovering a functionally meaningful structure of concept interactions. We next examine how much of this behavior is preserved by CONCEPTOME independently of the SAE reconstruction residual, addressing the reviewer's second concern.

 


W2: Contribution of CONCEPTOME and SAE reconstruction residuals

We thank the reviewer for suggesting these sanity checks. We agree that examining the contributions of CONCEPTOME-selected concepts and SAE reconstruction residuals provides a more complete understanding of the recovered structure.

Our original evaluation preserves SAE reconstruction residuals to isolate the effect of CONCEPTOME-based concept filtering from degradation introduced by SAE reconstruction. Importantly, our SAEs are trained independently on the representations of their respective layers, rather than to operate sequentially without their reconstruction residuals. Similarly, CONCORD learns dependencies between concepts at consecutive layer pairs without exposure to the effects of sequential residual removal. Discarding residuals throughout the network introduces cumulative reconstruction errors and distribution shifts affecting both the backbone's computation and the concept representations on which CONCORD operates. Preserving the residual therefore allows us to evaluate the discovered structure within the model it aims to explain, without introducing an additional source of degradation.

This choice is consistent with previous work in mechanistic interpretability. Marks et al. (2025) demonstrate that discarding SAE reconstruction errors can severely disrupt model performance, while Marks et al. (2025), Kim et al. (2025), and Chanin et al. (2025) account for reconstruction residuals when evaluating SAE-based explanations.

Nonetheless, following the reviewer's suggestion, we provide the requested global residual ablations, together with additional experiments to better contextualize their results. We first establish the reconstruction and functional fidelity of our individual SAEs, then assess the contribution of CONCEPTOME-selected concepts at each layer independently, before presenting the global ablations. This allows us to distinguish the contribution of the recovered concept structure from the cumulative effects of sequential residual removal.

1. Individual SAE fidelity. We first note that our SAEs achieve good reconstruction and functional fidelity when evaluated individually without their residuals. Across layers, ViT-B/32 achieves R² scores of 0.875 to 0.933 and classification accuracy of 58.1 to 62.0%, compared with 62.6% for the original backbone. For RN50, R² ranges from 0.845 to 0.993, with accuracy of 52.9 to 57.3% compared with 57.5% for the original backbone. We have included the complete reconstruction and functional fidelity results for each SAE in Appendix D.1 (Tables 4 and 5, p. 18) of the revised manuscript.

2. Independent-layer residual ablation. Motivated by the cumulative reconstruction errors and distribution shifts discussed above, we first evaluate CONCEPTOME without reconstruction residuals at each layer pair independently. At the intervened layer, we discard the reconstruction residual and compare retaining all SAE concepts against retaining only those selected by CONCEPTOME. In both conditions, the remaining downstream computation proceeds through the original backbone without further concept filtering. This allows us to isolate the contribution of CONCEPTOME-selected concepts at individual layers under identical downstream computational conditions, without accumulating the effects of sequential SAE reconstruction and concept filtering.

The following results are averaged across the 11 layer pairs of ViT-B/32 and the 15 layer pairs of RN50 (%).

ModelSubsetVariantAgreementFaithfulnessCompletenessDisappearance
ViT-B/32ImagenetteFull SAE without residual86.797.999.4-
CONCEPTOME without residual81.294.495.281.87
Residual only0.200.930.00-
ImageMeowFull SAE without residual89.498.496.8-
CONCEPTOME without residual86.397.096.579.70
Residual only0.233.230.00-
ImagewoofFull SAE without residual82.898.395.3-
CONCEPTOME without residual77.596.895.962.40
Residual only0.172.810.00-
RN50ImagenetteFull SAE without residual86.498.1100.0-
CONCEPTOME without residual86.197.8100.070.92
Residual only0.002.240.00-
ImageMeowFull SAE without residual89.498.899.7-
CONCEPTOME without residual88.698.499.771.78
Residual only0.000.730.00-
ImagewoofFull SAE without residual81.798.199.9-
CONCEPTOME without residual81.197.999.974.10
Residual only0.002.080.00-

Across both backbones and all three subsets, CONCEPTOME-selected concepts achieve 77.5–88.6% agreement and 94.4–98.4% faithfulness without residuals at the intervened layer, versus at most 0.23% and 3.23%, respectively, for residuals alone. These results show that the selected concepts preserve most of the predictive behavior supported by the full SAE representations without relying on residuals.

 


Reviewer_ssgUOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces CONCORD, which learns threshold rules between sparse autoencoder concepts in adjacent layers. These rules form a CONCEPTOME that can be inspected globally or queried for particular images and classes. Experiments on two CLIP backbones and three ImageNet subsets examine rule structure, prediction preservation under concept filtering, and the effects of concept removal.

Critical strengths

The graph shows concepts shared across classes, and removal tests support the importance of the selected concepts.

Critical weaknesses
  1. Table 2 does not test the symbolic surrogate on its own. It tests the original backbone after concept filtering and preserves the SAE reconstruction residual. The backbone still performs the computation. Prediction agreement for the learned rule chain alone is not reported, leaving the main surrogate claim untested.

  2. The rules' reliability is unclear. Table 1 reports tail consistency of 48.48-59.04%. This is the average fraction of tail concepts present when a rule fires, not how often the whole rule holds. The paper does not report complete-tail agreement or explain concrete rule failures.

  3. Training and evaluation use different interventions as written.

Questions for authors
  1. What prediction agreement does the learned rule chain achieve on its own, including its endpoint rules?

  2. For a representative rule, how often is its complete tail present, and do interventions that cross its activation threshold produce the predicted tail changes?

  3. Which masking and residual-handling procedures were used in training and evaluation, and how do they support the claimed match between the two?

2 comments

We thank the reviewer for their careful reading of our paper and constructive feedback. Below, we address the points raised and present additional analyses of the recovered dependencies.

 


Q1&W1: Standalone Prediction Agreement

We first clarify the objective of our work. The CONCEPTOME is intended as an interpretable structural surrogate of the model's concept-level organization. Our goal is to discover how concepts interact, combine, and are reused across samples and classes throughout a neural network, rather than to construct an alternative predictor. Unlike sample-specific circuit discovery, our approach learns a single reusable structure across samples and classes, taking a step toward the challenging goal of recovering a global, interpretable description of a modern vision model's internal organization. To this end, CONCORD is trained independently on consecutive layer pairs using pooled, binarized concept activations, abstracting away activation magnitudes and spatial locations. Given this abstraction, the composed rule chain, including its endpoint rules, is not expected to independently replicate the original model's computation. Accordingly, in Table 2 we evaluate the functional fidelity of the recovered structure rather than standalone prediction agreement, assessing how much of the original model's behavior is preserved or disrupted by retaining or suppressing concepts according to the CONCEPTOME. Notably, the CONCEPTOME achieves over 80% faithfulness across the evaluated architectures (CLIP ViT-B/32 and CLIP RN50) and all three datasets. On ViT-B/32, this is achieved while retaining only 73–75% of the active concepts. These results support the functional relevance of the recovered structure within the model it aims to explain. We have updated the manuscript to better reflect this distinction between a structural and a predictive surrogate.

 


Q2&W2: Rule consistency and interventional validation

Our objective is to discover recurring concept dependencies that collectively characterize the model's concept-level organization. As discussed in the Limitations section of the manuscript, the learned rules should not be interpreted as establishing every individual connection as a causal mechanism. CONCORD optimizes concept reconstruction across the entire rule bank, where multiple rules contribute to the prediction of a concept through OR aggregation, rather than enforcing complete-tail consistency for each individual rule.

Following the illustrative setting in Figures 4 and 5 of the original manuscript, we continue our analysis of CLIP ViT-B/32 on IMAGEMEOW. Beyond the average tail consistency reported in Table 1, we evaluate the probability of observing at least a fraction q of a rule's tail, conditional on the rule firing (Figure 10 of the revised manuscript, p. 24), with complete-tail consistency corresponding to q=1. Complete-tail consistency is 35.94% in the first layer pair and generally decreases with depth. In contrast, the probability of observing at least half of a rule's tail ranges from 58.60% to 98.15% across all layer pairs. These distributions show that rules continue to capture recurring subsets of their associated tail concepts throughout the network.

To assess the functional influence of these dependencies, we account for the fact that crossing the activation threshold of a single rule may leave other rules predicting the same target active. We therefore perform target-centric interventions on 100 target concepts per layer pair, evaluating each over the entire validation dataset. For each initially active target, we suppress the union of the antecedents of all fired rules predicting it and measure its disappearance rate. Figure 11 of the revised manuscript (p. 19) shows the resulting distributions, with layer-wise medians ranging from 9.86% to 19.28% and some targets disappearing in most of the samples in which they were initially active and predicted. These localized effects complement the 50.5% tail disappearance reported in Table 2 on IMAGEMEOW when intervening on the recovered structure collectively within each layer pair.

Complementing these local analyses, the original experiments in Figure 5 demonstrate that pathways composed of the recovered rules identify concepts with greater functional influence than matched random controls, particularly in later layers. This highlights the value of the CONCEPTOME in capturing computational organization beyond individual dependencies. We have incorporated these additional analyses of rule consistency and target-centric interventions into Appendix D.2 (p. 19) of the revised manuscript to provide a more comprehensive evaluation of the recovered rules and their functional influence.

 


Q3&W3: Training versus Evaluation

We use the same masking procedure during training and evaluation, replacing non-selected concept activations with their median values and preserving the SAE reconstruction residuals. Appendix A.5 did not reflect this implementation detail and has been updated accordingly.

Our SAEs are trained independently on their respective layers, rather than jointly to operate sequentially without reconstruction residuals. Discarding these residuals across layers can introduce cumulative reconstruction errors and distribution shifts that affect both the backbone and the learned rule banks. We therefore preserve the residuals to isolate the effect of the CONCEPTOME structure from that of the SAE reconstruction errors. However, CONCORD learns dependencies exclusively between SAE concepts, so that the recovered structure describes concept-level dependencies, while information carried by the residual remains outside this description. This residual preservation choice aligns with established practices in mechanistic interpretability: Marks et al. (2025) demonstrate that discarding SAE reconstruction residuals can disrupt model performance, and prior work, including RRM (Kim et al., 2025), Marks et al. (2025) and Chanin et al. (2025) account for these residuals during feature interventions.

Overall, discovering a single sparse, reusable, and functionally relevant concept graph across the layers of modern vision models represents a meaningful advance toward global mechanistic interpretability. Our results demonstrate that the CONCEPTOME captures shared computational structure across samples and classes while preserving much of the original model's computation. The additional analyses clarify both questions, and we have revised the manuscript to reflect these findings.

 


References

  • Chanin, D., et al. A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders. In NeurIPS, 2025.
  • Kim, J., et al. Interpreting Vision Transformers via Residual Replacement Model. In NeurIPS, 2025.
  • Marks, S., et al. Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models. In ICLR, 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.