acceptodds
Under review as a conference paper at ICLR 2027

AUTOREGRESSIVE MODELS REPRESENT MULTI-TOKEN NUMBERS AS PLACE-SCALED HELICES

Abstract

Large numbers are often split across multiple tokens by traditional tokenizers in modern Language Models (LM), raising a fundamental question: how are the numerical values represented when no single token corresponds to the entire number? Prior work reads the whole number from one hidden state, usually the last token. We show that this last token correlates with the number but does not causally control the other digits, and therefore in this work we analyse each token’s hidden state separately. We find that every token follows the same helical structure, with the period of the helix growing with the token’s place. We call this the Place-Scaled Helix. It correlates with the number representations better than prior structures, and intervening model activations using it changes the output digit, showing causal relevance. We have analysed the number representation across single-digit and multi-digit tokenizers, LLMs and VLMs, on arithmetic tasks showing a generalized numeric structure for autoregressive language models.

Primary area:
interpretability and explainable AI
Keywords:
Representation Learning, Mechanisitic Interpretability, Explainable AI

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_uGZ1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how autoregressive models represent numbers split across multiple tokens. It proposes a Place-Scaled Helix, where the period of each token’s helical representation scales with its numerical place. Experiments across LLMs, VLMs, and different tokenizers, together with activation interventions, support the proposed structure and its causal relevance.

Critical strengths
  1. Important research question. The paper extends the study of numerical representations from single-token to multi-token numbers, making the analysis more relevant to modern language models.
  2. Causal evaluation. Beyond measuring correlations, the authors intervene on model activations to test whether the proposed representations influence numerical outputs.
Critical weaknesses
  1. The scope of generalization needs clarification. The study focuses on arithmetic tasks, so it remains unclear whether the same structure applies to decimals, negative numbers, or numbers in broader language contexts.
  2. The causal interpretation needs more support. A last-token intervention failing to change other digits does not necessarily show that the last token has no causal role, especially if those digits were generated earlier.
  3. The manuscript lacks an AI Use Statement. The authors should add this section to disclose any use of AI tools, following the venue’s requirements.
  4. More related-work discussion. It would be better to discuss relevant work on interpreting arithmetic computation, e.g., ''Interpreting and Improving Large Language Models in Arithmetic Calculation'' and ''Case-based or Rule-based: How Do Transformers Do the Math?'' Comparing the findings with the proposed place-scaled representation would help clarify the paper’s novelty and its connection to arithmetic mechanisms.
Questions for authors

Refer to weakness.

3 comments
AuthorsOpenReview ↗

Rebuttal Response 1

(1) “scope of generalization needs clarification”:

We thank the reviewer for raising this point. As mentioned in Section 3.1 [Line 145], we analyze only natural numbers and extend prior work to multi-token numerical representations, which remain largely underexplored and are crucial. Other number settings are an important direction for future work, as acknowledged in Section 7 (Future Work)

(2) “causal interpretation needs more support”:

We apologize for the confusion. As the reviewer correctly points out, intervening on output tokens cannot affect digits that have already been generated. This is exactly why we perform causal steering on the input operand (a) and measure its effect on the subsequently generated output digits, as shown in Fig. 2 (right) and Table 5.

To further emphasize that we need to use all tokens instead of the last token alone, we performed linear probing on the last token to recover the entire number, but we can't steer all digits using the last token, as shown below (results are with Molmo2 in addition task)

10810^810710^710610^610510^510410^410310^310210^210110^110010^0
Probe accuracy: based on last token(%)10099.710010099.8100100100100
Steerability: digit changed (%)0.00.00.00.70.00.73.422.143.0

This is shown in Fig. 2 (Right): steering the last token in the operand only changes the corresponding digits (1 digit in the single-digit tokenizer and 3 digits in the multi-digit tokenizer). We can add 1 more digit if there’s a carryover.

(3) “manuscript lacks an AI Use Statement” :

We apologize for the omission and thank the reviewer for highlighting this requirement. We will add it in the revised manuscript. We generated and verified the evaluation datasets and carefully designed the experiments. AI was used only minimally to rephrase the manuscript text.

(4) “More related-work discussion”:

We thank the reviewer for suggesting these relevant works on arithmetic computation. Our work provides complementary representation-level explanations to their findings.

As shown in Fig. 4, numerical value representations become increasingly monotonic in later layers, coinciding with the emergence of the Place-Scaled Helix structure. Fig. 7 (bottom) shows high intervention accuracy (IIA) when the helix forms.

This explains the finding in ''Interpreting and Improving Large Language Models in Arithmetic Calculation'', where answers are processed in later layers, with earlier layers processing the operands and operator with the help of formation of place-scaled helix structures at respective layers.

AuthorsOpenReview ↗

Gentle Reminder

Thank you for your constructive feedback. We have addressed all the concerns raised in your review. We would appreciate it if you could review our rebuttal and consider revising your score if you find our concerns satisfactorily addressed. Just a gentle reminder that the Author-Reviewer discussion deadline is in 2 days (18th Nov AoE).

Thank you for the clarifications. I will keep my current score for two reasons. First, focusing only on natural numbers in arithmetic narrows the paper's overall contribution. Second, the related-work discussion remains somewhat brief and would benefit from a more thorough comparison with both suggested papers to better clarify this work's novelty.

Reviewer_9qAdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper analyses number representations in language models in the case of numbers being split across multiple tokens. The authors first show general properties of number representations by demonstrating that the representation's PC1 score changes linearly with magnitude and revealing distinct frequency components in the representations. Further, the authors find the representations of the output tokens to cluster according to place early in the network, motivating the number representation as a place-scaled helix. The place-scaled helix is then fitted to a low dimensional subspace of the output number representation, obtained by using Distributed Alignment Search. Activation patching is used to confirm that the helix number representations encode the relevant numeric information for prediction.

Critical strengths
  • The finding that the output number representations can be fit with a place-scaled helix is novel and interesting. In particular, the described sudden increase in quality of fit depicted in figure 7 constitutes an intriguing phenomenon.
  • Empirically, the number output representation of a place-scaled helix is shown to be rather general as it holds across different models and model sizes.
Critical weaknesses
  • The paper would overall benefit from clearer phrasing and more precise descriptions. Many sections do not clarify which residual streams are analysed and in which experimental settings the analysis is performed, making parts of the paper hard to read and understand. This is particularly the case for the introduction and section 3, where analysis of output and input representations seem to be mixed.
  • A more thorough analysis of the presented findings would strengthen the paper. While it shows that place-scaled helices are a good fit for the output representation in the last layers, it is not analysed how this structure arises, why it occurs particularly in addition tasks and why the quality of fit jumps suddenly around a specific layer.
Questions for authors
  • It is unclear, which representations are steered across section 3. The section first states to steer the residual stream of the operand "a" at different positions, but also steers the respective the output representations in the plots. Furthermore, the paper does not sufficiently differentiate between numerical input and output representations and does not clarify which are analysed in the respective sections. Can the place-scaled helix be found in both?
  • The introduction states that causal interventions on the last tokens do not affect the numeric behaviour. Could the authors clarify the meaning behind this? The later experiments show that representation manipulations of the operand do affect the prediction at the respective decimal place.
  • Why does the proposed representation structure appear to be task-specific? Does the bad fit for multiplication and division tasks result from the overall poor accuracy, or is the place-scaled helix representation mainly found in the output of subtraction and addition tasks?
  • The statements "correlation does not imply causation" and "emergence of causality" are not clear to me. Do these statements refer to the fact, that the representation of the place-scaled helix only appears to be a good fit only from a certain layer onwards?
  • In section 3.3 the different directions obtained from ridge regression are interpreted as varying numerical structure across decimal places. How can this be distinguished from interactions with the positional encoding and context in early layers?
5 comments
AuthorsOpenReview ↗

Rebuttal Response 1

(1) “which residual streams are analysed”:

We thank the reviewer for the suggestion. The input residuals (operand ‘a’) are used only for Fig. 2 and Fig. 3, to show how numbers are represented regardless of task or interventions. All other tables and figures are measured using output residuals to show how they depend on task.

As mentioned in Appendix A.4, for intervention experiments, we steered input operand residuals and measured changes in output number residuals, as shown in Table 5. For R2R^2 fit experiments, we also measured the output residuals to show the impact of the arithmetic task.

(2) “why the quality of fit jumps suddenly around a specific layer”:

We thank the reviewer for raising this point. From Fig 3, the helix representation is formed in earlier layers for the operand (input) and in later layers for the answer (output), which is correlated with the answer being formed in later layers in prior works [1,2,3].

[1] - Addition in Four Movements: Mapping Layer-wise Information Trajectories in LLMs

[2] - Arithmetic Without Algorithms: Language Models Solve Math with a Bag of Heuristics

[3] - A Mechanistic Interpretation of Arithmetic Reasoning in Language Models using Causal Mediation Analysis

(3) “why it occurs particularly in addition tasks”:

We have found that the representations for numbers are task-dependent. While the Place-Scaled Helix provides a strong fit for addition and subtraction, multiplication and division exhibit weaker fits and causal effects, suggesting that additional task-specific numerical structures may be involved. Understanding the underlying reasons for these differences is an interesting direction for future work. Our primary goal focuses on multi-token representation, which is largely underexplored and crucial. Further, the place-scaled helix shows the best fit overall compared with other complex fits.

Furthermore, we introduced the Free-period Helix, a place-scaled helix with periods chosen from the data rather than fixed (2,5,10,100). Interestingly, this formulation achieves improved fits for both addition and division, further supporting the relevance of the generalized helical structure.

Addition (9-digit answers)

Fit10810^810710^710610^610510^510410^410310^310210^210110^110010^0mean
Free-period Helix (9).87.67.69.67.65.65.66.65.82.70
Place-scaled helix (9).69.56.58.56.55.55.55.53.67.58
Polynomial (9).75.03-.01-.01-.01-.01-.01-.01-.01.08
Cubic B-spline (9).81.02-.01-.01-.01-.01-.01-.01-.01.09
Normal helix {100,10,5,2}\{100,10,5,2\} (9).17-.01-.01-.01-.01-.01-.01.30.67.12
Place-scaled circle T=10p+1T=10^{p+1} (3).42.28.31.30.29.30.30.28.31.31
Circle T=10T=10 (3).18.00.00.00.00-.01-.01.01.31.05

Division (6-digit quotients)

Fit10510^510410^410310^310210^210110^110010^0mean
Free-period helix (9).82.59.54.40.44.51.55
Place-scaled helix (9).67.45.44.31.34.38.43
Polynomial (9).70.01-.01-.01-.01-.01.11
Cubic B-spline (9).74.01-.01-.01-.01-.01.12
Normal helix {100,10,5,2}\{100,10,5,2\} (9).12-.01-
Place-scaled circle T=10p+1T=10^{p+1} (3).33.23.24.15.17.17.22
Circle T=10T=10 (3).13.00-.01.00-.01

(4) “Can the place-scaled helix be found in both input and output?”:

Yes, the place-scaled helix is found in both, as suggested in Fig. 3. In Fig. 3, we steered both the input and output to show that helix formation is universal for numbers in both. Although the degree of fit is task-dependent, that's why we analyze the output.

(5) “The section first states to steer the residual stream of the operand "a" at different positions, but also steers the respective the output”:

We apologize for the confusion. Our steering experiments are solely focused on steering the input operand ‘a’, because that's how we can show the change in all tokens in the output. Steering output tokens when they are already generated won't change the generated tokens. We will make this clear in the literature.

AuthorsOpenReview ↗

Rebuttal Response 2

(6) “Causal interventions on the last tokens do not affect the numeric behaviour. Could the authors clarify the meaning behind this?”:

We thank the reviewer for highlighting this ambiguity. As suggested by the steering experiment in Fig. 2 (Right), steering the last token only changes the digits corresponding to the last token, i.e., the units place (0–9) for single-digit tokenizers and the last three digits (0–999) for multi-digit tokenizers. Further, it doesn't affect other digits (except when there's carryover), even when it represents the entire number.

(7) “proposed representation structure appear to be task-specific?”:

We thank the reviewer for raising this question. Yes, the proposed representation is task-specific. The place-scaled helix is strongly correlated only for addition and subtraction. Although the place-scaled helix outperforms other fits, further studies are needed for multiplication and division, which is beyond the scope of this paper.

(8) “Does the bad fit for multiplication and division tasks result from the overall poor accuracy?”:

No, the task accuracy doesn’t impact the fit, all the results shown in the manuscript were with close to 100% accuracy. That's the sole reason we limited multiplication and division to 6 digits to maintain high accuracy.

We further measured the R2R^2 for the place-scaled helix in the multiplication task by varying the difference in the number of digits between the answer and the larger operand (d=0d = 0 to 33). We fix the larger operand aa to 3 digits and pick the smaller operand bb so that y=a×by = a \times b has exactly 3+d3+d digits, for d=0d = 0 to 33. For each dd we keep 50 distinct problems that the model answers fully correctly, so accuracy is 100% in every setting. We then fit the place-scaled helix at each answer digit and report R2R^2. (Empty cells (–) are the answer's leading digit, which we leave out as the value range is 1- 2, not the full 1-9 range).

The fit remains poor across all scales, suggesting that the representation depends on the arithmetic operation rather than the difference in numerical scale. We will clarify this in the revised manuscript. (The empty cells maintain the leading-digit distribution between 1-9, since we fix one operand to 3 digits.)

Placed=0d=0d=1d=1d=2d=2d=3d=3
units.62.54.63.56
tens.39.34.40.12
hundreds–.34.40.11
thousands––.22.27

(9) “"correlation does not imply causation" is not clear to me”:

We apologize for the confusion. It is mentioned because we can linear probe the entire number from the last token, but we can't steer all digits using the last token, as shown below (results are with Molmo2 in addition task)

10810^810710^710610^610510^510410^410310^310210^210110^110010^0
Probe accuracy: based on last token(%)10099.710010099.8100100100100
Steerability: digit changed (%)0.00.00.00.70.00.73.422.143.0

(10) “"emergence of causality" is not clear me”:

We apologize for the confusion. We show that place-scaled helix formation leads to better causal effectiveness during interventions in later layers (after Layer 23 for molmo2 and Layer 18 for Llama3.1), as shown in Fig. 7, and correlates with answers formed in later layers.

(11) “interactions with the positional encoding and context in early layers”:

We thank the reviewer for raising this point. Our interpretation is supported by causal interventions as shown in Fig. 2, right, where steering along these directions selectively changes the corresponding output digits. This shows that the directions capture relevant numerical information rather than merely encoding token positions.

AuthorsOpenReview ↗

Gentle Reminder

Thank you for your constructive feedback. We have addressed all the concerns raised in your review. We would appreciate it if you could review our rebuttal and consider revising your score if you find our concerns satisfactorily addressed. Just a gentle reminder that the Author-Reviewer discussion deadline is in 2 days (18th Nov AoE).

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

The paper studies numerical representations of multi-token numbers in LLMs and VLMs, and proposes a helical structure, motivated by the 3 observed properties: 1. Monotonicity of the representations 2. Their Fourier decompositions and 3. scaling along the digit position. The underlying structure hypothesis is then evaluated through fitting in a learned numerical subspace and various interventions across models, modalities and arithmetic tasks.

Critical strengths

The question of representations and their underlying geometry of numbers in LLMs and VLMs is rather interesting and important. In particular, the separation into one-token and multi-token classes of models is revealing in itself and invites for a separate study of each class. The paper presents results for a variety of models, both LLMs and VLMs, tokenizer types so the experimental setting is particularly diverse.

Critical weaknesses

My main concern is that the helical structure seems forced rather than proved in a meaningful way. A variety of geometries could satisfy the three observations motivating the proposed representation, and the experiments do not clearly distinguish the helix from other plausible structures. In particular, the proposed helix contains a linear component together with several sine/cosine pairs and is therefore considerably more expressive than the linear and logarithmic baselines. Its superior fit over these baselines is consequently not strong evidence for helicity.

Remarks:

Line 027: "has played an important role in advancing mathematical reasoning..." This does not seem right, perhaps supply with a reference or reformulate.

Figure 1 captions: It is not clear at this point why is n considered in the context of the addition task?

Section 3.2 As the authors already start presenting some of the motivating results here, it would be more reader friendly if there is a short description of the setup here. In particular, it is not clear what prompts are passed to the models in order to capture the intended numbers.

Line 155: Monotonicity, perhaps authors may wish to compare with AlQuabeh H. et al. Number representations in LLMs: A computational parallel to human perception, arxiv:2502.16147, as they studied similar phenomenon there. Also the claim that "the number representations are monotonic..." is too strong.

Before equation (1), "... we assume that the structure is helical", this is not clear. The authors proceed with Equation (1) from which it seems that the number representations are 3-dimensional, what is intended by this?

Line 197: "...proving that linearity exists" does not seem write. Perhaps approximate linearity or similar.

Line 200: This is step is not clear.

Line 203: Reads as "number representation has ... for a single-token number representation". It should be rephrased and at the same time the claim should be weaken, as it is not true written as is.

Equation (1): The set T\mathcal T is written weirdly, it should be simply be T={2,5,10,100}\mathcal T=\{2,5,10,100\}.

Section 4. Line 264: The claim that the place-scaled helix is the “minimal basis” satisfying the three listed constraints is not demonstrated. Many non-helical bases could satisfy monotonicity, periodicity, and place dependence.

Equation (3): Both here and equation (1) it is not clear what yy is. Whether it is a number, or a digit, and in particular what is intended dimension of the presented vectors.

Section 5. Line 315: The DAS training objective should be defined more explicitly. Since the subspace is trained to support numerical interventions, fitting and intervening with the helix inside it does not constitute fully independent causal evidence for the proposed geometry.

Tables 3–5: The candidate baselines have different dimensionalities and expressive power. Parameter-matched alternatives—such as polynomial, spline, random Fourier, and unrestricted Fourier bases—would provide a more informative comparison.

Questions for authors

See Weaknesses.

5 comments
AuthorsOpenReview ↗

Rebuttal Response 1

(1) “helical structure seems forced rather than proved”:

We thank the reviewer for raising this concern. Based on prior literature [1,2,3,4], numerical representations have shown circular structures with sine and cosine components (e.g., clock algorithm). With further analysis of linearity and periodicity, we hypothesized that the simplest structure combining circular and linear representations is a helix. We will clarify this motivation and distinction in the revised manuscript.

[1] - Language Models Encode Numbers Using Digit Representations in Base 10

[2] - Progress Measures for Grokking via Mechanistic Interpretability

[3] - Pre-trained Large Language Models Use Fourier Features to Compute Addition

[4] - Language Models Use Trigonometry to Do Addition

(2) “Reference for advancing mathematical reasoning”:

We thank the reviewer for pointing this out. Our intention was to highlight the growing use of LLMs for numerical and mathematical reasoning tasks. Therefore, we need to understand how these models internally represent and manipulate numerical quantities. We will cite prior works [1,2]

[1] - Language Models are Few-Shot Learners (Brown et al., 2020)

[2] - Training Verifiers to Solve Math Word Problems (Cobbe et al., 2021)

(3) “Figure 1 captions:”

We appreciate the reviewer's suggestion. We wanted to show the number representation for the operand, and the answer follows a helical structure. We will appropriately add the answer to show emphasis on the task.

(4) “a short description of the setup”:

We thank the reviewer for the suggestion. We describe the arithmetic evaluation setup with examples in Appendix A.4. We will make it clearer in Section 3.1 of the main paper.

(5) “perhaps authors may wish to compare with AlQuabeh H. et al.”:

We thank the reviewer for pointing out the relevant work. The numerical representations exhibit approximately monotonic ordering under the PC1 projection in our experimental setting. To validate this, we computed Spearman's rank correlation and observed (|ρ\rho| > 0.9) across the models and tasks, suggesting strong monotonicity. This observation is also consistent with the findings of AlQuabeh et al. We will clarify the scope of our claim and include the relevant citation in the revised manuscript.

(6) “The number representations are 3-dimensional, what is intended by this?”:

Apologies for not making it clear in the literature. To clarify, Equation (1), defines a 9-dimensional hyperhelix (1 linear, 4 sine, and 4 cosine basis functions) in the design matrix, rather than 3-dimensional. This formulation is suggested by our linearity and periodicity experiments in Fig. 2.

Once we have 9 scalar values for the given number 'n', we use a learned mapping from R9→R20\mathbb{R}^9 \rightarrow \mathbb{R}^{20}, which is embedded in the model's 4096-dimensional residual stream. This means linear combinations of these 9 basis-function scalar values should give us the 4096-dimensional vector for the number 'n'. We will clarify this distinction in the revised manuscript.

(7) “"...proving that linearity exists" does not seem write.”:

We thank the reviewer for the suggestion. As mentioned in [Line 158], we found approximate linear separability on PC 1 axis, showing approximate linearity as shown in Fig.2 left. We will revise the wording accordingly.

(8) Line 200: This is step is not clear:

We apologize for the confusion. We collect the residual stream activations across samples, perform PCA, and retain the top 32 principal components. For eahc component we try to fit sine and cosine functions using least squares at 600 different frequencies:

rd(v)≈adcos⁡(2πfv)+bdsin⁡(2πfv)for each f separatelyr_d(v) \approx a_d\cos(2\pi f v) + b_d\sin(2\pi f v) \quad \text{for each } f \text{ separately} And compute RSS amplitude across the 32 components , A(f)=∑d=132Ad(f)2,Ad(f)≈ad2+bd2A(f) = \sqrt{\sum_{d=1}^{32} A_d(f)^2}, \qquad A_d(f) \approx \sqrt{a_d^2 + b_d^2}

This helps us identify the dominant periodicities in the numerical representations, as shown in Fig.2 mid. We will make this clear in the Appendix

(9) “it is not clear what y is”:

The y is a scalar number as mentioned in Equation (1), and in Equation (3), in order to consider multi-token, we used a 9-digit number with a uniform distribution of digits for each place, as pointed out in Section 3.1 (Setup) and Appendix A.4. We will make this clearer in the revised manuscript.

AuthorsOpenReview ↗

Rebuttal Response 2

(10) “Parameter-matched alternatives”

We thank the reviewer for this valuable suggestion. We agree that only helix and place-scaled helix have 9 basis functions, while the remaining baselines have fewer. So, we have added higher-complexity fits with parameter-matched 9 functions. The table below shows R2R^2 for additional fits on Molmo2. Here, the place-scaled helix shows a better fit across digits.

Furthermore, we introduced the Free-period Helix, a place-scaled helix with periods chosen from the data rather than fixed (2,5,10,100). Interestingly, this formulation achieves improved fits for both addition and division, further supporting the relevance of the generalized helical structure.

Addition (9-digit answers)

Fit10810^810710^710610^610510^510410^410310^310210^210110^110010^0mean
Free-period Helix (9).87.67.69.67.65.65.66.65.82.70
Place-scaled helix (9).69.56.58.56.55.55.55.53.67.58
Polynomial (9).75.03-.01-.01-.01-.01-.01-.01-.01.08
Cubic B-spline (9).81.02-.01-.01-.01-.01-.01-.01-.01.09
Normal helix {100,10,5,2}\{100,10,5,2\} (9).17-.01-.01-.01-.01-.01-.01.30.67.12
Place-scaled circle T=10p+1T=10^{p+1} (3).42.28.31.30.29.30.30.28.31.31
Circle T=10T=10 (3).18.00.00.00.00-.01-.01.01.31.05

Division (6-digit quotients)

Fit10510^510410^410310^310210^210110^110010^0mean
Free-period helix (9).82.59.54.40.44.51.55
Place-scaled helix (9).67.45.44.31.34.38.43
Polynomial (9).70.01-.01-.01-.01-.01.11
Cubic B-spline (9).74.01-.01-.01-.01-.01.12
Normal helix {100,10,5,2}\{100,10,5,2\} (9).12-.01-
Place-scaled circle T=10p+1T=10^{p+1} (3).33.23.24.15.17.17.22
Circle T=10T=10 (3).13.00-.01.00-.01
AuthorsOpenReview ↗

Gentle Reminder

Thank you for your constructive feedback. We have addressed all the concerns raised in your review. We would appreciate it if you could review our rebuttal and consider revising your score if you find our concerns satisfactorily addressed. Just a gentle reminder that the Author-Reviewer discussion deadline is in 2 days (18th Nov AoE).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.