acceptodds
Under review as a conference paper at ICLR 2027

When Activation Patching Admits Many Faithful Circuits

Abstract

Activation patching can establish faithfulness without determining a unique minimum circuit. In every additive -component model with minimum -faithful cardinality , complete order- evidence admits at most minimum circuits. A patched ReLU model with patching sites attains the bound whenever and , giving an exact extremal characterization. For this construction, raising the patch-order bound to makes the full mask uniquely faithful and gives every proper mask faithfulness error at least one. At , the exact maximum is . For fixed , exponential multiplicity persists with per-component fan-in and parameter count bounded independently of and with uniform response norms through order . The extremal family and transition remain stable under component perturbations and bounded departures from additive masking. Exact noiseless deterministic identification of an unknown extremal circuit requires queries of order at most in the worst case for .

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, activation patching, circuit faithfulness, circuit identifiability, extremal combinatorics

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_d4nfOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper tackles a genuinely interesting question: how much can complete low-order activation patching evidence actually tell us about the uniqueness of a minimum faithful circuit? Working in an additive model with nn maskable components and minimum faithful cardinality kk, the authors bound the number of minimum faithful circuits by (n−1k){n-1\choose k}. What's particularly nice is that they then construct a patched ReLU that attains this bound, in this example, going from order rr to order r+1r+1 abruptly makes the full mask the uniquely faithful one. The paper doesn't stop there: it also provides a distributed construction with bounded local component complexity, tolerance-based stability results, an approximate-additivity extension, and a query lower bound for identifying an unknown member of the constructed candidate family. The validation here is purely mathematical, explicit constructions and proofs, no experiments on trained networks.

Critical strengths
  1. What I found most convincing is how the paper combines an upper bound, an attaining construction, and an explicit next-order transition. The ambiguity survives complete evidence on the continuous low-order domain, so it cannot be waved away as insufficient sampling, a weak search procedure, or an incomplete verification step. This is exactly why the paper's distinction between faithfulness and unique identification feels genuinely useful rather than purely terminological.

  2. The first construction concentrates a lot of cancellation in a single component, and honestly, without the distributed version I would have been much less convinced by the neural realization. By demonstrating exponential multiplicity with bounded local fan-in, parameter count, and response norm, the authors make it much harder to attribute the phenomenon to that particular modeling choice. That is a real strength of the paper.

Critical weaknesses
  1. My main hesitation concerns the gap between these additive constructions and ordinary circuit ablations in a network with downstream dependencies. The paper is well aware of this and provides an approximate-additivity theorem, which points in exactly the right direction. That said, asking for a small uniform residual over all relevant patches and all masks is quite a strong requirement, and the paper stops short of establishing when this condition actually holds in a representative non-additive setting. To be clear, I do not believe a theory paper needs a transformer benchmark to be publishable. Even so, a concrete account of where this assumption can be certified in practice would make a substantial difference to the significance of the result.

  2. Interestingly, the count (n−1k){n-1\choose k} can already be attained with constant additive components: simply take fj=1f_j=1 for j<nj<n and fn=−(n−1−k)f_n=-(n-1-k). In other words, the counting phenomenon itself is not specific to bounded-order activation patching. What feels more distinctive is that the paper realizes this count together with low-order equivalence and next-order separation. A related scope issue arises with robustness. Exact cancellation is generally destroyed by arbitrarily small perturbations, so the stability results understandably work with a positive tolerance, though the allowable per-component perturbations typically shrink as nn grows. The formal statements are mostly careful about this, but I would like to see the high-level discussion held to the same standard. Let me stress that these are limitations on the breadth of the contribution, not counterexamples to the stated theorems.

Questions for authors
  1. Could you spell out how the cited Móricz–Nagy theorem and lemmas yield the integer zero-sum statement used in Lemma 23, including any relevant restrictions on their applicability? The universal bound rests on this step, and I would like that dependency to be readily auditable.

  2. Does the p=2qp=2q case of the distributed construction also have the full mask as its unique order-(r+1)(r+1) faithful mask? My reading of the witness equations is that they force each block to be selected as a whole, and then matching the full response forces all blocks to be retained. If this is correct, I think it deserves an explicit corollary, since it would nicely strengthen the bounded-complexity part of the paper.

  3. Is there a standard masking setting, beyond independent parallel branches or prescribed readout-edge groups, in which the required uniform additive-surrogate residual can already be bounded? An existing analytic example would really help me judge whether the approximate-additivity extension is mainly a continuity statement around the construction, or a useful route to less idealized circuit analyses.

4 comments

Thank you for your detailed and constructive review of our submission.

Q1. Zero-sum bound. For Lemma 23 we use Theorem 1.7 of Móricz and Nagy (arXiv:2411.04449v2). Remove the component responses that vanish on the allowed patch domain. Let NN be the number remaining and qq the maximum number whose sum vanishes there. Appendix C.4 constructs a linear functional on their rational span that takes rational values and is nonzero on every nonzero subset sum. Applying it and clearing denominators gives nonzero integers a1,…,aNa_1,\ldots,a_N with exactly the original zero-sum subsets. Let bi=i+2 aib_i=i+\sqrt{2}\,a_i for 1≤i≤N1\leq i\leq N. Each bib_i is irrational, and equality bi=bjb_i=b_j forces ai=aja_i=a_j and then i=ji=j. For every qq-element index set TT, the sum ∑i∈Tbi\sum_{i\in T}b_i is rational exactly when ∑i∈Tai=0\sum_{i\in T}a_i=0. For N/2<q<NN/2<q<N, Theorem 1.7 bounds the number of such subsets by (N−1q−1)\binom{N-1}{q-1}. The reduction therefore needs no separate appeal to their Lemmas 2.1 and 2.2. For 1≤q≤N/21\leq q\leq N/2, we apply the Erdős–Ko–Rado theorem. Two disjoint maximum zero-sum omissions would have a larger zero-sum union. The family is therefore intersecting. If q=0q=0, the empty set is the only possible omission among the remaining components. Your question led us to clarify both ranges after Theorem 11 (Section 5, p. 6). Appendix C.4 gives the reduction in the proof of Lemma 23 (pp. 18–19).

Q2. Paired blocks. Yes. In the paired-block construction of Theorem 14, let qq be the number of blocks and p=2qp=2q the number of positive components. For the patch-order bound rr in that theorem, the full mask is uniquely faithful at order r+1r+1. The new corollary follows your argument. Each block has two positive components and one cancellation component. Moving from the zero patch to the order-(r+1)(r+1) witness patch for positive component jj changes the masked score by the retention indicator of jj minus that of its cancellation component. The full score stays at qq. Faithfulness forces these indicators to agree for every jj. Each block must therefore be retained or omitted as a whole. Since each retained block contributes one, matching qq requires all qq blocks. Every proper mask misses the full score by a nonzero integer at zero or at a witness. Omitting one whole block gives error exactly one everywhere, proving the gap is sharp (Section 5, corollary after Theorem 14, p. 7, and Appendix C.6, pp. 20–21).

Q3. Residual branches. Veit et al. (2016, Section 4.1) delete individual residual branches while retaining the identity connections. We analyze all four masks on two consecutive branches under this operation. Let hh be the patched input, u,vu,v the branches, ww the linear readout vector, and x1,x2x_1,x_2 the binary retention indicators of the two branches. Let the additive surrogate match the empty and singleton masks. The masked score minus the surrogate score is x1x2w⊤[v(h+u(h))−v(h)]x_1x_2 w^\top[v(h+u(h))-v(h)]. If vv has Euclidean Lipschitz constant at most LvL_v and ∥u(h)∥2≤U\|u(h)\|_2\leq U on all allowed patches, the absolute residual is at most ∥w∥2LvU\|w\|_2 L_v U for every mask and patch. We derive a bound from the weights for ReLU branches and an exact value for linear branches with affine patched input (Section 6, p. 9, and Appendix D.1.1, pp. 24–25).

W1. Residual certificate. Your suggestion prompted the scalar example. For h=z∈[0,1]h=z\in[0,1], w=1w=1, u(h)=h/8u(h)=h/8, and v(h)=h/10v(h)=h/10, the uniform residual is 1/801/80. The larger surrogate singleton error is 10/8010/80, and the surrogate empty-mask error is 18/8018/80. For each mask, Theorem 19 bounds the change in error by 2/802/80. At tolerance 14/8014/80, both singletons are faithful (error at most 12/8012/80), and the empty mask is unfaithful (error at least 16/8016/80). These bounds certify precisely the two singleton masks as minimum in a model where deleting the first branch changes the input to the second (Section 6, p. 9, lines 471–476).

W2. Counting and stability. Your constant-component example now follows Proposition 9. For nn components and minimum faithful size 1≤k<n1\leq k<n, its (n−1k)\binom{n-1}{k} minimum circuits agree at every patch order. With mm patching sites, 0≤r<m0\le r<m, and n−1≤(mr+1)n-1\le\binom{m}{r+1}, the patched ReLU construction in Theorem 10 attains the same count through order rr and separates every pair at order r+1r+1 (Section 5, pp. 5–6, and Section 8, p. 10).

As you note, arbitrarily small perturbations can destroy exact cancellation. The abstract and Section 1 now specify positive tolerance for stability. At tolerance 0<ε<10<\varepsilon<1, Theorem 17 preserves the constructed family when the component perturbations satisfy its uniform bound and Equation (3). For fixed ε\varepsilon and k=⌈n/2⌉k=\lceil n/2\rceil, the bound on the uniform perturbation per component decreases as 1/n1/n. The unperturbed next-order gap of one is 1/k1/k of the full score (Section 5, pp. 7–8, and Section 8, p. 10).

We would be grateful for any further suggestions you may have on the revision.

Thank you for adding the residual example. In Section 6, it establishes two minimum faithful circuits at positive tolerance, but does not demonstrate exponential multiplicity or the uniqueness transition at order r+1r+1.

In Theorems 10 and 14, the component responses are constant on Zr\mathcal{Z}_r, so faithfulness reduces to one scalar constraint. These constructions are valid, but their relevance to broader nonadditive networks remains unclear. Please keep this limitation explicit when discussing the residual example.

Overall, I keep weak reject recommendation but reduce my conf. to 2.

Thank you again for your careful review and further comments.

We agree that the scalar example alone does not establish exponential multiplicity or the transition at order r+1r+1. Theorem 19 and equation ~4 already establish their preservation under the stated residual bound. Following your comment, we have added a corollary to the existing Theorem 19 that provides an explicit residual network satisfying that bound.

  • Multiplicity and transition. For a positive integer qq, the corollary constructs 3q3q consecutive residual branches from qq paired blocks. Under the stated patch-site assumptions, exactly (2qq)\binom{2q}{q} masks of size qq are minimum faithful at patch order rr and tolerance 1/21/2. At order r+1r+1, the full mask is uniquely faithful at the same tolerance. The residual relative to the paired-block surrogate is uniformly at most 5/325/32 over all masks and patches of order at most r+1r+1, satisfying equation (4). The corollary appears in Section 6, page 9, with its proof in Appendix D.1.2, pages 25–26.

  • Nonconstant responses. Your observation about the constructions in Theorems 10 and 14 is correct. In the added construction, every singleton response varies on the low-order patch domain. At an allowed patch, deleting the first branch changes the contribution of the second. Each branch uses O(r+1)O(r+1) nonzero local parameters, independently of qq. These conclusions concern the explicit constructed family under the stated assumptions. The scalar example retains its separate role as a certificate for two consecutive branches.

We accommodated the corollary within ten pages by shortening the residual calculations in Section 6 and consolidating few details in Appendix D.1.1. The existing mathematical results and their numbering remain unchanged.

We would be grateful for any further suggestions you may have on the revision.

Thank you for the additional construction. The new corollary in Section 6 and its proof in Appendix D.1.2 address my main remaining concerns. They establish exponential multiplicity and the transition at order r+1r+1 in a nonadditive residual network with nonconstant singleton responses. I therefore raise my rating to Weak Accept. The lack of experiments on trained networks remains a limitation on practical relevance.

Reviewer_dsNpOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper asks whether activation patching can uniquely identify the components that implement a model behavior. It studies additive component models and complete patching evidence up to order r, meaning that every allowed intervention changing at most r activation sites is considered. The authors prove that, for n components and minimum faithful circuit size k, the number of minimum circuits consistent with this evidence is at most ((n-1) over k). They show this bound is tight by constructing patched ReLU models in which that many distinct minimum circuits agree on every patch of order at most r. In the construction, patches of order r+1 can distinguish the candidate circuits, and every proper subset of the full model has a faithfulness error of at least one. The paper also presents a construction with bounded per-component complexity, studies stability under small perturbations and departures from additivity, and gives a lower bound on the number of queries needed to identify an unknown circuit. Validation is theoretical: the claims are established through proofs and constructed examples, without experiments on real-world models.

Critical strengths

The paper develops a potentially interesting theoretical perspective on the identifiability limits of activation patching. Its sharp upper bound, together with a construction that attains it, gives a precise account of how many minimum faithful circuits can remain indistinguishable under bounded-order evidence. This could help clarify what activation-patching results can and cannot establish, and supports the paper’s value as a theoretical contribution.

Critical weaknesses

The paper lacks clear empirical experiments on realistic models. Its main claims are supported by mathematical proofs and constructed examples, so it remains unclear how often the identified ambiguity arises in practical activation-patching analyses. This limits the evidence for the paper’s broader relevance to mechanistic interpretability.

Additionally, the proofs are generally rigorous and readable for a mathematically trained audience, but the main argument is technically involved and relies on several reductions to combinatorial results. A clearer roadmap and more intuition around these reductions would improve accessibility.

Questions for authors

Do the authors see practical applications of their results to activation-patching analyses of real models? In particular, can the proposed bounds or constructions help practitioners assess whether a circuit identified from bounded-order patching evidence is unique?

Do similar ambiguities arise in practical activation-patching analyses of real-world models? If so, can the authors provide evidence or examples showing how often multiple minimum faithful circuits remain indistinguishable under bounded-order patching?

2 comments

Thank you for your thoughtful review of our submission.

More than one minimum mask may pass a finite patch test. Your question about testing uniqueness led us to clarify in Section 7 when the finite-set test also establishes uniqueness on the full patch domain.

Q1. Testing uniqueness. For an additive score, choose a finite patch set and a faithfulness tolerance. First, minimize the number of retained components subject to matching the full score within tolerance on every patch in the set. Then, restrict the search to masks of optimal size, exclude the first mask, and solve again. With rational data and exact optimization, a second solution exhibits two minimum masks, and infeasibility certifies uniqueness on the tested set. If we also certify that the unique optimum is faithful on the entire continuous domain at the same tolerance, it is uniquely minimum there as well. Any competing mask of equal or smaller size that is faithful on the entire domain is also feasible on the tested set (Section 7, Proposition 20, p. 10, and Appendix D.2.1).

The example in Section 7 shows why the last check matters. At tolerance 14/8014/80, testing only the zero patch gives the empty mask as the unique minimum. Testing only the patch at one gives two singleton optima. For each mask, the score difference from the full model is linear in the patch value and vanishes at zero. Its maximum absolute value on [0,1][0,1] occurs at one. Both singleton masks are therefore minimum on the whole interval (Section 7, final paragraph, p. 10).

Q2. Trained models. Your question about trained models prompted our rebuttal pilot test. We used ten trained digit classifiers and the same ten held-out image pairs for each classifier. We enumerated all 256 readout masks for each model–pair case. For the digit-0 logit minus the digit-1 logit, each of the 100 model–pair cases had a unique minimum under exact agreement on the complete continuous order-one patch domain for that pair. On each one-site interpolation, checking the endpoints and ReLU breakpoints suffices because the score difference between each mask and the full model is affine between consecutive breakpoints. The supplementary archive contains the code, trained weights, and results for each case in the work_exp/ directory.

W1. Empirical scope. We agree that the constructions do not establish how often trained models admit multiple minimum faithful masks under exact agreement. The 100 pilot cases cover one architecture and one score. Méloux et al. (2025, Section 4.3) found 3,209 circuits satisfying their behavioral criterion in the final layers of one network trained on MNIST digits 0 and 1. Their count establishes multiplicity under that criterion. It leaves open whether two circuits both have minimum size and reproduce the full score on every allowed patch (Section 8, final paragraph, p. 10).

W2. Proof accessibility. Following your suggestion for a clearer guide to results and proof reduction, we replaced the main-text proof sketches with high-level explanations and intuitions. Section 5 now starts the counting argument with the components a faithful circuit omits. Under additive masking, the full score minus the masked score is the sum of the omitted component responses. Exact faithfulness means the sum vanishes on every allowed patch. Minimizing the retained size therefore maximizes the number omitted. After removing components whose responses vanish on every allowed patch, two nonempty maximum zero-sum omissions must intersect, since otherwise their union would be a larger zero-sum omission. The intersecting-family bound applies when a maximum omission is nonempty and contains at most half of the remaining components. For larger omissions, the integer zero-sum bound supplies the count. We replaced the main-text proof sketches for Proposition 9 and Theorem 11 with these explanations (Section 5, pp. 5–6). Appendix C opens with a guide to the argument (p. 15, lines 800–803), and Appendices C.2 and C.4 give the detailed derivations (pp. 16 and 18–19).

We would be grateful for any further suggestions you may have on the revision.

Reviewer_dsNpOpenReview ↗

Borderline paper, but raising the recomenndation to weak accept

Thank you for the detailed responses and revisions. I still consider this a borderline paper: the theoretical results are interesting, but their practical applicability to circuit analysis in realistic models remains uncertain. Nevertheless, the revisions improve the presentation and clarify the scope and limitations of the results, addressing several of my concerns. On balance, I am raising my recommendation to weak accept.

Reviewer_RxgpOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies whether complete bounded-order activation-patching evidence determines a unique minimum-cardinality faithful circuit. For an additive model with nn maskable components and minimum exactly faithful circuit size kk, it proves that the number of minimum faithful circuits is at most (n−1k)\binom{n-1}{k}. The proof relates faithful circuits to complements of maximum-cardinality zero-sum omission sets and combines an intersecting-family argument with an existing extremal zero-sum result.

The authors construct a patched ReLU model attaining this bound on the complete continuous order-rr intervention domain. Increasing the allowed intervention order to r+1r+1 makes every proper mask unfaithful. Additional results include a construction with bounded per-component complexity, stability guarantees under positive-tolerance faithfulness and approximate additivity, a finite-domain uniqueness test, and a sharp deterministic query bound for identifying an unknown candidate through its response oracle. Validation is theoretical, through explicit constructions and proofs; the paper does not report trained-model experiments.

Critical strengths
  1. The paper gives a sharp bound on the number of minimum exactly faithful circuits, together with a matching ReLU construction. This provides a precise quantitative characterization of circuit non-uniqueness.

  2. Exponential ambiguity persists over the complete continuous low-order intervention domain and with bounded component complexity. This demonstrates a structural limitation of the admitted patching evidence.

Critical weaknesses
  1. The additive masking assumption substantially restricts applicability. The approximate extension requires a uniform residual bound over all masks and patches, which the paper does not establish for a relevant trained model.

  2. The sharp count concerns exact faithfulness, while the robustness results preserve a particular family under positive tolerance. Moreover, the guaranteed component perturbation radius shrinks as O(1/n)O(1/n), and the normalized next-order gap is 1/k1/k. These limitations weaken the practical interpretation of the robustness claims.

Questions for authors
  1. In which intended applications can the additive assumption or the required uniform residual bound be justified?

  2. How should the robustness and next-order transition be interpreted given the shrinking perturbation radius and normalized separation?

  3. What circuit-discovery setting provides access to the unknown-candidate response oracle assumed in Theorem 13?

2 comments

Thank you for your careful review of our submission.

As you point out, the next-order gap depends on the scale of the score. Section 5 now gives the exact threshold for relative error, and Section 6 derives the residual of an additive surrogate for two consecutive residual branches.

Q1. Additive masking. Masking disjoint groups of edges entering a linear readout gives exact additivity. At a fixed patch zz, deleting a group removes its contribution to the score while preserving the inputs to the other groups. Additivity holds for learned readout weights and for activations that depend nonlinearly on zz (Section 5, p. 4).

For two consecutive residual branches, deleting the first changes the input to the second. For linear branches with weight matrices AA and BB, patched input h(z)h(z), and readout vector ww, retaining both branches adds the interaction w⊤BAh(z)w^\top BA h(z). The additive surrogate agrees with the masked score on the other three masks. With mm patching sites and order bound 0≤s≤m0\le s\le m, the domain Zs\mathcal{Z}_{s} consists of the vectors in [0,1]m[0,1]^m with at most ss nonzero coordinates. The exact uniform residual is ηs=sup⁡z∈Zs∣w⊤BAh(z)∣\eta_s=\sup_{z\in\mathcal{Z}_s}|w^\top BAh(z)|. For an affine patched input, sorting coefficients obtained from the weights and the clean and corrupted activations gives this supremum (Section 6, p. 9, and Appendix D.1.1, p. 24).

Q2. Radius and gap. In the sharp construction of Theorem 10, let nn be the number of components, rr the patch-order bound, and kk the minimum faithful size at order rr, with 1≤k<n1\le k<n. Adding the same constant δ>0\delta>0 to each of the n−kn-k responses omitted by a minimum circuit gives that circuit an error of (n−k)δ(n-k)\delta at order rr. Uniform preservation at tolerance 0<ε<10<\varepsilon<1 therefore requires δ≤ε/(n−k)\delta\leq\varepsilon/(n-k). As you observed, this upper bound decreases as 1/n1/n for fixed ε\varepsilon and k=⌈n/2⌉k=\lceil n/2\rceil. In the unperturbed construction the smallest order-(r+1)(r+1) error of a proper mask is exactly one and the full score is kk. With error measured relative to the full score, the full mask is uniquely faithful at order r+1r+1 exactly when the relative tolerance τ\tau satisfies 0≤τ<1/k0\le\tau<1/k (Section 5, discussion after Theorem 17, p. 8, and Appendix C.9).

Q3. Oracle access. The analyst knows the model and the minimum size kk. Each query specifies a patch and returns the score of one fixed, hidden minimum mask from the construction in Theorem 10. In our construction each patch of order at most r+1r+1 tests membership of at most one of n−1n-1 components. The known size determines the last membership once the other n−2n-2 are known. Theorem 13 proves that n−2n-2 noiseless queries are necessary in the deterministic worst case and sufficient for 1≤k≤n−21\leq k\leq n-2. Your question prompted us to clarify that in ordinary circuit discovery the analyst chooses the mask to evaluate (Section 5, Theorem 13, pp. 6–7, and Appendix E).

W1. Applicability. We agree that the paper does not verify the required residual bound and margins for a trained network with downstream dependencies. Your comment prompted us to show how to verify both in a model with two consecutive residual branches. In the scalar example, the largest absolute residual over all masks and allowed patches is 1/801/80. At tolerance 14/8014/80, Theorem 19 gives an error of at most 12/8012/80 for each singleton and at least 16/8016/80 for the empty mask. Thus, the two singletons are precisely the minimum faithful circuits (Section 6, pp. 8–9).

W2. Positive tolerance. You are right that the sharp upper bound (n−1k)\binom{n-1}{k} in Theorem 11 concerns exact faithfulness in an additive model. Arbitrarily small perturbations can destroy exact cancellation. At positive tolerance, all (nk)\binom{n}{k} size-kk masks can be minimum in an additive model. Theorem 17 gives sufficient conditions for the sharp construction. If each component response changes uniformly by at most δ\delta through order r+1r+1 and Equation (3) holds, the same masks remain minimum at order rr and the full mask remains uniquely faithful at order r+1r+1, both at tolerance 0<ε<10<\varepsilon<1 (Section 5, pp. 6 and 8).

We would be grateful for any further suggestions you may have on the revision.

Thank you for the detailed response.

The examples clarify where additivity holds, and the followup discussion answers my questions. My main reservation remains the connection to practical circuit discovery. I appreciate the theoretical characterization. I will raise my score to weak accept

Comments on the paper

1 comment

We sincerely thank all reviewers and the Area Chair for their time, constructive feedback, and continued engagement. We also appreciate the reviewers’ willingness to raise their ratings/scores and their decisions to do so. Your incredibly constructive reviews have significantly improved our submission.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.