acceptodds
Under review as a conference paper at ICLR 2027

Causal Separation: A Theory of Representation and Recursive Self-Improvement

Abstract

Recursive self-improvement aims to let machines improve how they learn from experience. A system can continue to act on a familiar relation after the conditions supporting it have changed. The resulting failure exposes a dependence that sustained its activity without being represented. We characterize this problem by tracing the closed causal paths through which activity continues. An initial separation between represented and unrepresented relations leaves participating support on one side and its symbolic projection on the other. We show how this causal separation has corresponding environmental and representational forms. When the same represented state admits different consequences, we prove that their received separation cannot be generated by the old internal relations alone. A cycle of representational reproduction incorporates these symbolic effects into reusable relations and changes subsequent interaction. We derive its coupled response and conditions for learning without labels for hidden causes. In finite settings, the cycle recovers after support changes and removes error floors that further fitting within the old representation cannot overcome. For a two-layer linear family, the original gradient update rule executes the self-modifications selected from experience. With consistent estimation, selection, and execution, the system asymptotically achieves the lowest attainable expected loss after a fixed training period.

Primary area:
learning theory
Keywords:
recursive self-improvement, representation learning, learning dynamics, causal modeling, learnability

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_uuKdOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies how a learning system can identify relations that its current representation cannot express and use those relations to modify its subsequent learning process. It introduces “causal separation” to connect unrepresented conditions affecting the system with their observable consequences. In finite settings, the paper develops partition-refinement and response-feature constructions that remove error floors imposed by a fixed representation. For a two-layer linear network, it characterizes function-restoring parameter preparations executable through the original gradient flow using internally generated targets, and analyzes how experience can guide their selection to improve subsequent adaptation. A synthetic experiment examines the roles of source correspondence, relation formation, and deployment.

Critical strengths
  1. The paper usefully distinguishes limitations in information availability, representational capacity, and deployment. These are different reasons why additional experience may fail to improve a system. The formal analyses and targeted controls help make this distinction concrete.

  2. The constructive reachability result provides a concrete technical contribution. Theorem 19 characterizes the attainable preparations in the two-layer linear setting and constructs paths through the original gradient update rule.

Critical weaknesses
  1. The additional contribution of the unifying framework over established theory remains insufficiently clear. Several components are closely connected to existing ideas. The partition-refinement construction relates to state equivalence and model minimization in MDPs, as studied by Givan et al. [1]. Changing subsequent optimization behavior through parameter transformations also relates to the parameter-symmetry and teleportation literature, including Zhao et al. [2]. The authors acknowledge these connections, which is helpful. Nevertheless, I found it difficult to distinguish applications of established tools from the principal new theoretical advances, and to identify what additional conclusions specifically follow from the causal-separation framework. Theorem 19 appears particularly important in this respect: the paper would benefit from a more explicit comparison explaining what its reachability characterization establishes beyond prior parameter-transformation results.

  2. The connection between the concrete results and recursive self-improvement remains limited and highly conditional. The central neural construction concerns a particular two-layer linear family, internally controllable training targets, and a fixed subsequent adaptation period. The recursive analysis in Appendix F.7 assumes adequate candidate coverage, accurate evaluation, effective selection, and reliable execution at each generation. It remains unclear which concrete constructions preserve these conditions through successive self-modifications, and what the recursive result adds beyond a general accumulation of per-round approximation errors. The experiment in Appendix E.7 supports the proposed mechanism, but uses a tightly specified setting with a known observation alphabet, noise rate, future query, and gradient dynamics for candidate evaluation. These constrain how strongly the paper supports its broader contribution concerning recursive self-improvement. I do not consider large-scale application benchmarks necessary for this theoretical work, but a sharper account of the recursive guarantee and its significance would be more consequential.

[1] Equivalence Notions and Model Minimization in Markov Decision Processes.

[2] Improving Convergence and Generalization Using Parameter Symmetries.

Questions for authors

See weaknesses.

2 comments

Thank you for your careful review. We appreciate your recognition of the information, representation and deployment distinctions, and of Theorem 19. Your comments help us clarify the learning contribution and its recursive realization.

W1: The additional contribution of the unifying framework over established theory remains insufficiently clear. [...]

The contribution is a common prototype of learning as the reproduction of causal separation. A judgment becomes independent of particular training circumstances, while feedback continues to change the connections supporting it. Comparing their different consequences produces new representations through the use of existing ones.

This viewpoint is analogous to renormalization. Its four sectors are causal functions: a uses a judgment; b receives consequences; c compares them; d reconstructs the distinctions used next. Let KiK_i be their transition laws, with retained memory included. For coarse maps qiq_i preserving the next-interface law,

\quad\Longrightarrow\quad \bar T=\bar K_d\bar K_c\bar K_b\bar K_a.$$ Successive conditioning proves the implication. Different implementations can therefore share an effective learning cycle. Feedback failure, insufficient comparison capacity and failed renewed use remain distinct causes of similar plateaus. This is a supplementary formalization of the renormalization-like view. Givan et al. [1] supply the state-equivalence foundation. Our construction learns and rebuilds the grouping from encounters. Different consequences force distinctions; these can expose further differences in paths reaching a condition. Theorem 3 and Corollary 4 prove formation and recovery under specified sampling conditions. Zhao et al. [2] study parameter transformations and optimization benefits. We ask whether the learner can execute the change through its original gradient rule. Theorem 19 gives exactly $(QD,0)$, $Q\in SO(n)$, as the returned endpoints from $(D,0)$, with $D$ positive diagonal. The invariant $W^\top W-VV^\top=D^2$ excludes some equivalent-output points, including $(2D,0)$; the construction reaches every permitted rotation. The learning conclusion is that restoring the present prediction need not restore the old learning response. Experience can change how the same learner adapts next. Theorem 20 analyzes attainable successor selection under its information and control conditions. > W2: *The connection between the concrete results and recursive self-improvement remains limited and highly conditional. [...]* For sustained recursive learning, two requirements matter: the concrete return must remain executable, and its effective cycle must allow new distinctions to change later interaction. We now implement both in one continuous learning program. Neural parameters, source-bound records and reference directions persist. Returned observations guide gradient preparation; the actual trained output proposes a new environmental query; response comparisons validate it; execution generates the next data. The environmental operator is hidden, and later queries are generated internally. The existing neural construction supports repeated execution. With $r=Wv$, the invariant keeps $K=WW^\top+\\|v\\|^2I$ positive definite. From output $r_0$, internal error $-K^{-1}r_0$ gives $\dot r=-r_0$, returning to a state where preparation applies again. Thus this program carries the execution condition into the next round. The test completes 48 runs of 16 rounds. In 8D, response-law MSE is 0.00010, versus 0.75000 when new queries cannot change the environment. A paired same-encounter replay also measures a learning benefit from preparation. Regarding the information assumptions, the new program keeps a finite interface and the known internal update rule. It estimates variability from samples, generates its own queries, and prepares before the fresh signed target is revealed. F.7 identifies the coverage, evaluation, selection and execution errors that must remain controlled. Our finite-candidate experiment does not establish vanishing errors over the full rotation orbit. It tests a concrete continuing implementation. Progress additionally requires informative opportunities; external shutdown ends the return. These conditions explain what sustained self-improvement must preserve, rather than treating repeated execution alone as sufficient. Thank you again for encouraging this connection to a concrete recursive program. **References** [1] Givan, R., Dean, T., and Greig, M. Equivalence notions and model minimization in Markov decision processes. *Artificial Intelligence*, 147(1–2):163–223, 2003. [2] Zhao, B., Gower, R. M., Walters, R., and Yu, R. Improving Convergence and Generalization Using Parameter Symmetries. *International Conference on Learning Representations (ICLR)*, 2024.

Thank you for your thoughtful response and additional clarifications. My first concern has been largely addressed. Regarding my second concern, I remain somewhat uncertain about how the required conditions are maintained across successive self-modifications.

I must also acknowledge that my expertise in the relevant theoretical areas is limited, and I have already informed the AC of this. After considering your response and the other reviewers’ assessments, I have decided to maintain my current rating for now. Thank you again for your efforts.

Reviewer_oPLTOpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The manuscript seems to aim to develop a framework for understanding how a learning system can handle novel situations that cannot be explained by its current representations. It presents several theoretical results concerning causal separation and representation formation during learning. A two-layer linear neural network is used to illustrate some of these ideas.

I must admit that this topic is outside my area of expertise, and I had considerable difficulty understanding most parts of the manuscript. In particular, I found the paper extremely difficult to read and follow. Many important terms, such as "causal structure," "causal dependence," and "symbolic projection," are introduced without sufficiently clear explanations of their precise meanings or their roles in the proposed framework. As a result, I could not clearly understand the main contributions or how the different theoretical results support the central claims.

The authors acknowledge using generative AI for manuscript drafting and revision. While the use of AI is not itself a concern, I am worried that the manuscript has not been sufficiently revised to make its arguments clear and understandable to readers. The presentation often feels overly complicated and lacks clear explanations of the main ideas.

For these reasons, I currently assign a score of 1. However, given my limited background in this topic, I am open to reconsidering my assessment if the authors can clarify the main ideas and contributions during the rebuttal.

Critical strengths

see paper summary

Critical weaknesses

see paper summary

Questions for authors

see paper summary

2 comments

Thank you for the time and effort you invested in reviewing our paper. We appreciate your comments on the clarity of its central ideas and contributions. We would like to state our position more directly.

Motivation. In ordinary learning, the learner, training data and update process belong to one causal whole. Nevertheless, successful learning produces a judgment that can remain effective when particular examples, backgrounds or other training circumstances change. Our research takes a perspective analogous to renormalization: looking past the individual events, we identify the relative independence that has formed within this connected process. This is causal separation. External causes still affect the learner, but they no longer dictate its judgment through the same particular dependencies. For example, varied handwriting can establish a digit criterion that remains useful beyond the writers and images that produced it.

Developmental canalization provides a precedent: internal organization stabilizes outcomes across variations in underlying conditions [1]. A useful geometric picture is a vortex drawing diverse causal influences inwards. Their retained effects are compressed into reusable distinctions that preserve differences relevant to prediction and action. Our causal account connects this organization to learning. Encounters form internal distinctions that govern subsequent interaction; feedback can then reconstruct those distinctions and the connections supporting them. This is the recursive aspect of canalization: the structure formed through interaction becomes a constraint on further interaction, while remaining open to revision. Causal separation captures the relative independence that emerges through this process.

Approach. Our causal analysis connects this stability to an internal separation between symbolic and connective representations. The symbolic side expresses a definite distinction the system adopts; the connective side includes the relations through which that distinction works in concrete situations, still partly determined by the outside world. The boundary is therefore active rather than impermeable. The same symbol can persist while its connections encounter different consequences. Comparing those differences can produce new relations and new symbolic distinctions. This is why representational reproduction is a learning process: reuse becomes a means of forming representations that were not previously available.

Results and implications. This account also gives direction to RSI. At the concrete level, the learning return must remain executable: action must reach the environment, feedback must return, and comparison must be able to change subsequent activity. At the effective level, the separation must be continually reproduced through adoption, reception, comparison and reconstruction. The aim is a system that keeps developing its own judgment through interaction, including interaction that changes how it learns. The manuscript's finite constructions and neural self-modification results establish specific realizations of these requirements.

Overall, we believe this viewpoint connects familiar learning behavior, the formation of representations and the problem of sustained self-improvement. Thank you again for your careful reading and comments.

References

[1] Mather, K. Genetical control of stability in development. Heredity, 7:297–336, 1953.

I would like to thank the authors for their response and the attempts to clarify their paper. I have to say that I get lost starting from the 1st point on motivation. There are just many jargons like “causal whole”, “renormalization”, “developmental canalization” etc. As I mentioned in my original review, this difficulty may reflect my limited expertise in the field. For this reason, I assigned a confidence score of 1 to my review, which I hope the Area Chair will take into account. Nevertheless, based on my current understanding of the paper and the authors’ response, I do not have sufficient grounds to change my score.

Reviewer_6ZCMOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies recursive self-improvement: how a system can use experience to change the way it learns. The authors describe causal separation between represented relations and the unrepresented processes that affect a system's activity. They propose a cycle that uses observed consequences to form new relations and change later behavior. In finite settings, they prove recovery after support changes and removal of error floors caused by a restricted representation. For a two-layer linear family, they show that the original gradient update rule can carry out selected self-modifications. Under consistent estimation, selection, and execution, expected loss after a fixed training period approaches the minimum within the allowed modification class.

Critical strengths

The paper asks an interesting question: can a system form new representations from experience and use them to improve how it learns next? It connects representation learning with changes to the system's later learning behavior.

Critical weaknesses
  1. Recursive self-improvement is not directly tested. The experiment shows one round of modification and later training, but does not show whether the modified system can retain useful information, repeat the process, and improve again.

  2. Stable-partition learning and recovery after support changes are not tested. Theorem 3 and Corollary 4 provide theoretical guarantees, but the numerical study does not evaluate these procedures. The paper does not show experimentally how much data they need or how quickly they recover.

  3. The experimental setting is very limited. It uses a two-dimensional linear model with only one missing feature direction, so the choice of a new relation is heavily constrained. The results do not show whether the method can identify useful relations when several possibilities are available.

Questions for authors
  1. Could the authors show several rounds of self-modification? Can the modified system continue to collect useful information, form new relations, and improve again?

  2. Could the authors test stable-partition learning and recovery after a support change, as described in Theorem 3 and Corollary 4? How much data and time are needed to form correct partitions and recover after a change?

  3. Could the authors test higher-dimensional models or settings with several possible missing relations? Can the method identify which relations are useful? Could the authors also report the computation needed for candidate evaluation and preparation, and compare methods under the same total budget?

3 comments

Thank you for your detailed review. We appreciate your recognition of the connection between representation learning and later learning behavior. Your suggestions motivated further tests of repeated learning, reconstruction and cost. Our responses follow your questions.

Q1/W1: Could the authors show several rounds of self-modification? Can the modified system continue to collect useful information, form new relations, and improve again?

Thank you for identifying the need to test the complete return. We have now run a continuous neural–environment learning loop, joining operations previously tested separately.

One program retains its neural parameters, records and reference directions. Returned observations guide preparation through the original gradient rule. The network then trains on a new target, and its actual output proposes another query direction. Response comparisons validate it. The new query changes later evidence.

We ran 4D/8D systems, eight seeds and three conditions, for 16 rounds each: 48 runs. Every run uses 3,072 response vectors, noise SD 0.05 and adaptation time 0.04 per round. The hidden environmental law is not given to the learner. Complete neural states are inherited rather than replaced by prepared endpoints. Prediction error uses a ridge readout of the retained source–return records.

Control runs retain new distinctions but keep queries within the initial space. Their first two rounds match the full runs exactly.

Final response-law MSEFull loopNew queries blocked
4D0.000030.50001
8D0.000100.75000

New distinctions support learning by changing subsequent interaction. The law is also learned without directed preparation: the necessary ingredient in this contrast is access to new conditions, not this particular controller.

A paired replay audit compares preparation and no preparation from the same returned state on the same recorded encounter. In 8D, preparation lowers fixed-interval target loss by 0.044 [0.036, 0.052] (95% interval across eight seeds). This isolates its local learning benefit despite different subsequent query trajectories.

Q2/W2: Could the authors test stable-partition learning and recovery after a support change, as described in Theorem 3 and Corollary 4? How much data and time are needed to form correct partitions and recover after a change?

We tested the submitted rebuilding rule on eight records and three actions, with 20 seeds. The learner receives outcomes and transitions without the true partition or change time. All seeds recover the correct four-block partition at these first successful tested checkpoints:

StageSamples per source–action rowTotal samples
Formation32,768786,432
Recovery262,144 new6,291,456 new

The change follows 100,000 old samples per row, which remain in memory. Rebuilding removes obsolete distinctions; split-only refinement retains eight blocks. The unchanged initial grouping has value regret 0.6811, versus zero after recovery. Reconstruction takes about 0.20 ms at these checkpoints. Sampling uses vectorized controlled-row draws, so this is computation time, not physical acquisition time. Old counts and conservative thresholds slow recovery.

Q3/W3(a): Could the authors test higher-dimensional models or settings with several possible missing relations? Can the method identify which relations are useful?

The new 4D/8D systems can form two/six additional independent query directions beyond their initial references. Each round compares five/nine preparations: the current orientation plus four/eight candidate rotations. New query directions come from trained outputs and response comparisons. The paired replay in Q1 tests whether the selected preparation helps on the same encounter.

Q3/W3(b): Could the authors also report the computation needed for candidate evaluation and preparation, and compare methods under the same total budget?

The new loop scores the current orientation plus four/eight candidates, without training each candidate. Preparation averages 19,842/66,277 gradient-field calls and the same number of dense solves per round in 4D/8D.

Our separate 12-round budget study gives the following mean losses at matched gradient-call allowances:

DimensionPreparedDirect updates
4D0.3970.358
8D0.3340.205

Direct updates perform better. They spend the allowance after target revelation without the 0.04 deadline. Observation, scoring and dense-solve costs are not equalized. Thus this measures a gradient-call budget, not equal total FLOPs; preparation demonstrates a changed learning response, not a computational-efficiency advantage.

Thank you again for these concrete suggestions. They helped us test repeated execution and further acquisition within one continuing program.

Thank the authors for the replies and new experiments. In my original review, my main concern was that the paper presented extensive theoretical arguments with limited empirical evidence. The experiments provided in the rebuttal partially address my concerns.

Q1: The multi-round experiments address my concern about continued learning. However, as the authors acknowledge, the environment law can also be learned without preparation. If I do not misunderstand, this improvement cannot be fully attributed to self-modification.

Q2: The new experiments address my questions about partition formation and recovery after a support change. The reported sample counts and reconstruction time help clarify the practical cost of these procedures.

Q3: The 4D/8D experiments with multiple candidate directions address much of my scope concern. The budget comparison is also helpful. However, as the authors acknowledge, direct updates perform better under matched gradient-call allowances. My original concern about whether the benefits justify the preparation cost therefore remains.

I appreciate the authors' effort to provide this additional evidence. These experiments are helpful to strengthen the original submission. However, I think this paper needs further revision to integrate these results, clarify the limitations and polish the language before publication. So, I keep my original rating.

Thank you very much for your careful review and follow-up response. We realize there remain some misunderstandings regarding our work, and we believe the following clarifications will resolve your concerns.

Q1: The multi-round experiments address my concern about continued learning. [...] this improvement cannot be fully attributed to self-modification.

We would like to clarify preparation's relationship to learning in general and self-modification. Causal separation is our account of learning itself, including ordinary learning without directed preparation. Self-modification concerns changes in how subsequent learning proceeds; preparation is one construction through which the manuscript studies such changes. In the multi-round comparison, both groups learn. The contrast concerns what preparation adds to that shared process.

A network, its optimizer and its data participate in one causal process. Across training history, the retained effects of different examples constrain one another. A basis for judgment can emerge that works beyond the particular circumstances that produced it. This is causal separation: an open system develops relative independence within the process that shapes it. The manuscript's neural construction makes one consequence precise: identical current predictions can support different subsequent learning responses.

Q2: The new experiments address my questions [...] help clarify the practical cost of these procedures.

We are glad the supplementary experiments resolved your questions about partition formation and recovery, and clarified their cost. We will include the procedures and results in the revised manuscript.

Q3: The 4D/8D experiments [...] My original concern about whether the benefits justify the preparation cost therefore remains.

Thank you for following up. Our claim concerns causal separation as a learning mechanism; preparation is a probe. Obstructing this separation can impair later learning despite correct predictions on familiar inputs.

In the existing 4D/8D environment family, we wrote a copy rule into f(x)=WVxf(x)=WVx by setting W=V=IW=V=I. Familiar inputs lay in the environment's fixed-point direction, where copying is correct. The comparison network started at W=I,V=0W=I,V=0 and learned those answers through feedback. Both received 2,048 noisy SGD updates on familiar inputs, then identical broader inputs and feedback. The environment law, loss and optimizer stayed fixed; all parameters remained trainable.

After familiar training, both had MSE near 10−610^{-6}. Yet in 8D, reaching unfamiliar-input MSE below 0.05 required:

Initial condition, followed by familiar trainingUpdates, mean ± SD
No supplied answer218 ± 1
Supplied copy rule1,832 ± 11
Same copy, W=1.01I,V=I/1.01W=1.01I,V=I/1.01819 ± 5

Eight paired seeds; 4D gave the same ordering. The third condition preserves all initial predictions and changes the total squared parameter norm by only 0.02%.

The supplied rule preserved familiar success while obstructing the necessary new judgment. On a new input requiring -1, it initially answered +1. Feedback suppressed that answer toward zero, where it stalled long after the comparison network had learned -1. Correction reached the parameters, but initially worked mainly to erase the old answer.

On this response component, let a=wva=wv and c=w2−v2c=w^2-v^2. In the population dynamics, the manuscript's gradient equations give da/dt=−∣c∣da/dt=-|c| at a=0a=0. The supplied symmetric rule has c=0c=0: as correction suppresses the old response, it also extinguishes the capacity to cross zero. A nonzero cc permits the crossing and the formation of the required negative response. The rescaled control retains the supplied answer while releasing this bottleneck.

Familiar practice preserves this pattern: omitting that stage gives corresponding update counts of 218, 1,834 and 819. In the actual noisy SGD runs, symmetry eventually breaks and all groups recover within 4,000 updates.

Learning is how a system shaped by its background comes into its own. Accumulated experience must form a judgment that extends beyond its original support and remains open to revision. There is a kind of selfhood in this emergence. Our renormalization-like perspective follows this separation across training history. Symbolic distinctions and their environmentally dependent connections give it representational form; a–d describes their reproduction.

RSI requires this formation to sustain itself: further experience must keep renewing both judgment and learning. Inputs and gradients must do more than circulate through the machinery: their effects must become an enduring, revisable basis for the learner's own activity.

Thank you again for your time and continued engagement. We hope these clarifications and additional tests address your concerns, and we would be grateful if you might reconsider your evaluation in light of them.

Comments on the paper

1 comment

Dear Area Chair and Reviewers 6ZCM, uuKd, and oPLT,

Thank you for the time and effort you have invested in reviewing our paper. We particularly appreciate the recognition of the connection between representation learning and subsequent learning behavior, the distinction between information, representation and deployment, and the constructive contribution of Theorem 19. We believe the discussion, clarifications and supplementary experiments have substantially strengthened the work. Together, they make the paper’s central contribution clearer: a unified causal account of learning, representation formation and recursive self-improvement. We have responded to each review separately and summarize the main points below.

Reviewer 6ZCM recognized the connection between representation learning and later learning behavior, and asked for repeated self-modification, tests of stable-partition reconstruction, higher-dimensional settings and computational accounting. We have provided these tests and added a continuous neural–environment program in which the same trained neural state helps generate later queries and receives the consequences of those queries. The reconstruction and budget studies are reported alongside it, including the substantial sampling requirement and the cost comparison favoring direct updates.

Reviewer uuKd recognized the distinction between information availability, representational capacity and deployment, as well as the constructive contribution of Theorem 19. Their central concerns were the additional value of the unified framework and the conditions supporting recursive use. Our response explains the learning meaning of the reachability result and gives a compact four-sector account of representational reproduction. It then connects those requirements to a concrete continuing program, with its information, execution and continuation conditions stated explicitly.

Reviewer oPLT emphasized the difficulty of understanding the main ideas and the roles of the terminology. Our response starts from ordinary learning: a learner and its training environment form one causal process, but a learned judgment can continue to work when the particular circumstances that formed it change. This is the phenomenon that causal separation is intended to explain.

From a perspective analogous to renormalization, causal separation is the emergence of relative autonomy within a connected causal process. Its inward, compressive dynamics consolidate the effects of diverse encounters into learned distinctions that govern subsequent interaction. Within the learner, this separation takes representational form: stable symbolic distinctions coexist with connective relations still shaped by the environment.

RSI requires the continued reproduction of this separation. Adoption, reception, comparison and reconstruction turn further interaction into new representations. The causal loop must remain executable: action must reach the environment, feedback must return, and newly formed distinctions must change subsequent activity. Learning thereby renews both its representations and the conditions of further learning.

We thank the reviewers again for helping us make this position and its supporting evidence clearer. Detailed answers are provided under the respective reviews, and we welcome further discussion of any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.