acceptodds
Under review as a conference paper at ICLR 2027

MoE-Spec: Expert Budgeting for MoE Speculative Decoding

Abstract

Speculative decoding accelerates Large Language Model (LLM) inference by verifying multiple candidate tokens in parallel. For Mixture-of-Experts (MoE) models, each candidate token routes to different experts, and the target model must load the union of all activated experts to verify them. As the number of candidates grows, unique experts per layer approach the full model, and verification becomes as expensive as running a dense model. We propose MoE-Spec, a training-free method that enforces a fixed expert budget at each layer during verification. MoE-Spec aggregates routing probabilities across candidate tokens and loads only the top- experts. Because the remaining experts are excluded, the budget controls a tradeoff between throughput and quality. At conservative budgets, MoE-Spec matches EAGLE quality and often improves throughput; tighter budgets trade quality for additional throughput. Experiments on OLMoE-1B-7B, Qwen3-30B-A3B, and Mixtral-8x7B across reasoning, code generation, and summarization show that at conservative budgets MoE-Spec matches EAGLE quality while improving throughput, and at aggressive budgets retaining 90% of autoregressive quality, throughput improves 15–45% depending on model architecture. Code will be released upon acceptance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
llm acceleration, speculative decoding, mixture of experts models

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_cjiGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the expansion of the expert union during tree-based speculative decoding for Mixture-of-Experts (MoE) models, which increases expert-weight memory traffic during verification. It proposes MoE-Spec, which aggregates routing probabilities across the draft tree, selects a budgeted subset of experts at each layer, and adjusts token-to-expert assignments accordingly. The expert budget controls the trade-off between task quality and throughput. Experiments cover three MoE architectures and five benchmarks, reporting quality-matched speedups at conservative budgets and quality–throughput trade-offs at more aggressive budgets.

Critical strengths
  • The problem is clearly identified and supported by empirical analysis. The paper highlights the overhead caused by the growing union of activated experts during tree-based verification. Analyses of expert activation counts, draft-tree size, decoding speed, and runtime breakdown support this motivation.
  • The method is simple and requires no additional training. Expert selection uses existing routing probabilities, with an explicit budget controlling the quality–throughput trade-off. The method integrates into the existing EAGLE verification pipeline. At conservative budgets, several configurations achieve higher throughput without reducing the reported task metrics relative to EAGLE.
Critical weaknesses
  • Quality-matched efficiency comparisons against simple approximate inference baselines are missing. The authors report quality-matched speedups over EAGLE and compare the task quality of Static and Router selection in the ablation study. However, the experiments do not provide quality–throughput curves for simple alternatives such as EAGLE with a statically pruned expert set. Consequently, the additional benefit of dynamic, tree-level expert selection over these alternatives remains insufficiently established. End-to-end throughput should be compared at matched task quality under the same hardware and evaluation settings. Autoregressive decoding with expert pruning could serve as an additional baseline.
  • The routing concentration analysis contains some discrepancies between the figure and the text. The caption of Figure 7 states that MBPP achieves 95% routing probability coverage at budget B=8, and the accompanying text claims that the top eight experts capture over 90% of routing probability mass for both MBPP and HumanEval. However, the corresponding bars indicate approximately 65%, respectively. Similarly, the CNN/DM bar at B=24 indicates approximately, below the stated 90% (also, the bar chart is not suitable for this situation where we need to judge the numbers). These discrepancies weaken the routing concentration analysis and its explanation of why small expert budgets are viable. The authors should clarify the coverage definition and aggregation procedure, and reconcile the plotted results with the caption and text.
Questions for authors
  1. Could the authors provide quality–throughput curves comparing the proposed method with EAGLE? Including autoregressive decoding with expert pruning would further help isolate the efficiency gains attributable to dynamic, tree-level expert selection.
  2. The author should make it clear about routing concentration analysis.
2 comments

Thank you for the detailed review.

"Quality-matched efficiency comparisons against simple approximate inference baselines are missing ... Autoregressive decoding with expert pruning could serve as an additional baseline."

MoE-Spec is faster than both simple baselines at similar or better accuracy (Table C1) because neither chooses experts for the tree we verify. We ran EAGLE-3 with reduced top-k and with a statically pruned expert set and applied both only during verification on the same tree (Appendix F of the revision). Reduced top-k decides for each token alone and its drops cost accuracy but save no memory traffic when other tokens in the tree still load the expert. Static pruning fixes one set in advance and fails on MATH500 and the code tasks.

Table C1 Speedup over EAGLE-3 at batch 1 on the same tree. Parentheses give the accuracy change in points.

MethodGSM8KMATH500HumanEvalMBPPCNN/DM
MoE-Spec (32 experts per layer)1.16x (-1.0)1.20x (-0.7)1.19x (+0.0)1.20x (-0.7)1.09x (+0.0)
MoE-Spec (44 experts per layer)1.07x (-0.4)1.07x (-0.5)1.06x (+0.4)1.07x (+0.3)1.01x (-0.1)
Reduced top-k (4 of 8 per token)1.07x (-3.6)1.04x (-4.6)1.11x (-1.8)1.06x (-6.6)1.01x (-0.8)
Reduced top-k (6 of 8 per token)1.02x (-0.4)1.01x (-2.6)1.02x (-0.6)1.03x (+0.6)1.01x (-0.1)
Static pruning (64 of 128 per layer)1.07x (-2.5)1.02x (-11.6)1.07x (-28.7)0.93x (-41.0)0.97x (-1.3)

Autoregressive decoding with half of the experts of each token stays slower than EAGLE-3 on the reasoning and code tasks and loses up to 8.8 points (Appendix F).

Appendix F also plots quality against speed for every setting we measured. Neither baseline exceeds 1.11x there on any task while MoE-Spec reaches 1.20x at its aggressive budget for an accuracy loss of about 1 point.

"The routing concentration analysis contains some discrepancies between the figure and the text."

Thank you for spotting this issue. The bars of Figure 7 are correct but our caption and text mischaracterize them. We define coverage as the share of routing probability that the B experts with the most probability capture when we sum probability over the tokens of a draft tree. On code and math the top 8 of 64 experts capture nearly two thirds of the routing probability and the top 24 capture over 90%. On GSM8K and CNN/DM the top 8 capture about half and the top 24 capture over 85%. We replace the bars with a table of these values and correct the text in the revision (Appendix D).

Thanks for your clarification, and I will maintain my positive assessment.

Reviewer_MioUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper observes that in tree-based speculative decoding for MoE models, the union of experts activated by all draft-tree tokens quickly approaches the full expert set, so verification loses MoE's sparsity advantage and large trees become counter-productive. So, this paper proposes MoE-Spec, a training-free method that sums router probabilities across the draft tree and executes only the top-B experts per layer during verification, either truncating or substituting the out-of-budget experts for each token.

Critical strengths
  1. Clear and well-motivated observation. The "expert explosion" in MoE tree verification and the fact that EAGLE-3's speedup peaks and then degrades with tree size on MoE are useful, concrete diagnostics for the community.
  2. The method is training-free and needs only router outputs that are already computed, which is easy to understand and follow.
Critical weaknesses
  1. The method is lossy, but there are no lossy baselines. Once out-of-budget experts are dropped, the verifier is no longer the target model, so MoE-Spec is dynamic expert pruning rather than lossless speculative decoding. The paper never compares it with the same budget applied to plain AR decoding, so I cannot tell whether the gains come from the speculative setting or simply from running fewer experts.
  2. The EAGLE baseline is not tuned for MoE. All main results use a 60-token tree, yet the paper’s Fig. 3b shows EAGLE-3 on OLMoE peaks at about 31 tokens. At EAGLE’s best tree size, the gap is only 1.95× vs. 2.1×, much smaller than in Table 1. Since MoE-Spec gains most from oversized trees, this setup systematically favors the proposed method.
  3. Operating points are selected on the test set. The C and A budgets are picked per model and benchmark by inspecting test-set quality. A deployed system cannot pick B per task with oracle knowledge of quality. A fixed B, or a rule calibrated on held-out data, is needed to support the claims.
  4. No variance or significance testing.
  5. The claim about scaling to larger MoEs is untested. The paper repeatedly argues that DeepSeek-V3 or Qwen3-235B-scale fine-grained MoEs would benefit more. However, no such model is evaluated.
Questions for authors

Can you select a single B per model (or a rule such as a routing-mass threshold) on a held-out calibration set and report test results with that fixed choice, together with confidence intervals or multiple seeds for the quality metrics?

3 comments

Thank you for the detailed review.

"The method is lossy, but there are no lossy baselines ... I cannot tell whether the gains come from the speculative setting or simply from running fewer experts."

MoE-Spec is indeed lossy and its gains come from the speculative setting. As Table M1 shows, running fewer experts without a draft tree stays far slower than EAGLE-3 on the reasoning and code tasks.

Table M1 Speedup over autoregressive decoding at batch 1. Parentheses give the accuracy change in points against lossless decoding.

MethodGSM8KMATH500HumanEvalMBPP
Autoregressive (6 of 8 experts per token)1.07x (-1.4)1.07x (-1.8)1.12x (+0.6)1.11x (+0.8)
Autoregressive (4 of 8 experts per token)1.16x (-6.1)1.14x (-8.8)1.16x (-1.2)1.17x (-2.6)
EAGLE-3 (lossless)1.50x1.60x1.52x1.56x
EAGLE-3 + MoE-Spec (44 experts per layer)1.60x (-0.4)1.71x (-0.5)1.61x (+0.4)1.67x (+0.3)

At batch 1 an autoregressive step processes one token that loads only its own 8 experts per layer. Running fewer experts there means lowering k as in the first two rows. MoE-Spec instead removes experts that a draft tree loads but its tokens barely use. Appendix F of the revision also compares MoE-Spec with reduced top-k and static pruning during verification.

"The EAGLE baseline is not tuned for MoE."

On Qwen3-30B in SGLang, tuning picks an EAGLE-3 tree of about the same size as the default tree of Table 1. On this tree MoE-Spec is still faster on every task at the conservative budget (Table G1).

"Can you select a single B per model ... on a held-out calibration set and report test results with that fixed choice, together with confidence intervals or multiple seeds for the quality metrics?"

Yes, this is the protocol of the shared response. Table G1 reports its test results with paired 95% bootstrap intervals (Appendix F).

"The claim about scaling to larger MoEs is untested."

We agree that this claim is untested at that scale. As a proxy that fits on our hardware and needs no draft model, we compare Qwen3-30B-A3B with the sparser Qwen3-Next-80B-A3B. Both models read the same text from the five benchmarks and we count the experts that 64 consecutive tokens select together against one token (Appendix F). Verifying such a window loads 9x the experts of decoding one token on Qwen3-30B and 15x on Qwen3-Next. Because MoE-Spec removes exactly this excess, we expect larger gains on sparser models and state this as a hypothesis that the measurement supports.

Thank you for the clarifications. May I ask where Table M1 and Table G1 are? I suppose M1 is the table you showed here, which looks good. However, I can't find the G1 you are referring to.

Please see the shared response above for Table G1. We tried to unify questions we felt multiple reviewers asked in one place. After review, please let us know if you have any follow-up questions.

Reviewer_MD6pOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MoE-Spec, a training-free expert budgeting method for speculative decoding in Mixture-of-Experts (MoE) models. The authors identify a sparsity collapse in standard MoE speculative decoding: as the draft tree grows, the target model must load almost all experts to verify the tokens, negating the memory bandwidth benefits of sparse activation. MoE-Spec addresses this by aggregating routing probabilities across the draft tree and loading only the top-B experts per layer, redirecting tokens to available alternatives if their preferred experts fall outside this budget. Experiments on models like OLMoE, Qwen3, and Mixtral show that MoE-Spec achieves 15–45% throughput improvement over EAGLE while retaining ≥90% of autoregressive quality.

Critical strengths
  1. The paper identifies a structural conflict between tree-based speculative decoding and sparse MoE execution. As the draft tree grows, the union of activated experts approaches the full expert set. This is a clear and practically relevant systems bottleneck.

  2. The proposed MoE-Spec is training-free, requires no changes to draft construction, and elegantly controls the quality-throughput tradeoff via a single budget parameter.

Critical weaknesses
  1. The evaluation focuses exclusively on single-request inference. In batched serving scenarios where expert reuse across requests inherently changes verification cost, the relative throughput benefits of MoE-Spec might be diluted.

  2. Each benchmark contains only 80 evaluated samples. For accuracy-based metrics, the number of samples are insufficient.

  3. For each model–benchmark combination, the paper selects the budget that maximizes throughput while keeping quality within one standard deviation of EAGLE. It means it uses post-hoc budget selection on the test benchmarks

Questions for authors
  1. If the draft-tree size is tuned separately for EAGLE and MoE-Spec, what is the best throughput achieved by each method? How much of the reported 10%–30% advantage remains?

  2. Do the throughput gains persist at batch sizes 4, 8, and 16 ?

  3. Could the authors select BB on an independent validation/calibration split, freeze it, and then report quality and throughput on a held-out test split that was not used for budget selection?

1 comment

Thank you for the detailed review. We note that two points in it describe an earlier preprint of this work (80 samples per benchmark and budget selection within one standard deviation of EAGLE). The submission evaluates the full test sets (Section 4.1) and reports the iso-quality and 90%-of-AR points of Table 1.

"If the draft-tree size is tuned separately for EAGLE and MoE-Spec, what is the best throughput achieved by each method?"

Tuned in SGLang, EAGLE-3 runs fastest on a 64-token tree. On this tree MoE-Spec is about 7% faster than EAGLE-3 on the reasoning and code tasks at the conservative budget and up to 20% faster at the aggressive budget (Table G1). Larger trees accept more tokens per pass but run slower for both methods on all tasks but one.

"Do the throughput gains persist at batch sizes 4, 8, and 16?"

Up to batch 8 the gains persist and MoE-Spec is at least 6% faster on the reasoning and code tasks (Table G2 and Appendix F). At batch 16 the gain shrinks as the review anticipated and the conservative budget roughly breaks even. The aggressive budget is still faster there and loses less than 1 point.

"Could the authors select (B) on an independent validation/calibration split, freeze it, and then report quality and throughput on a held-out test split that was not used for budget selection?"

Yes, this is the protocol of the shared response. Table G1 reports its test results with paired 95% bootstrap intervals (Appendix F).

Reviewer_qt2kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies expert activation growth during tree-based speculative decoding for MoE models. MoE-Spec aggregates router probabilities across the verification tree, selects a expert subset at each layer, and restricts token routing to that subset. The method is training-free and explicitly trades output quality for reduced verification cost. Experiments cover OLMoE-1B-7B, Qwen3-30B-A3B, and Mixtral-8x7B on five benchmarks, with additional analyses of selection strategies, coverage policies, and reconstruction error. The reported results suggest useful quality–throughput operating points, although the quality-preservation protocol and several routing-related details require clarification.

Critical strengths
  1. A simple fix for a key inference bottleneck. Restricting expert unions directly addresses expert-weight traffic instead of relying solely on draft acceptance. It requires no retraining, offers an explicit resource-control parameter, and is highly useful for single-request inference.
  2. Consistent gains across models and tasks. Table 1 shows matched-score speedups, e.g., Qwen3/MATH500 (2.1× to 2.5× at acc 71) and Mixtral/GSM8K (1.6× to 2.0× at acc 44). Ablations and runtime profiling further support these findings beyond raw speedup numbers.
Critical weaknesses
  1. Quality-constrained speedups lack robustness guarantees. It is unclear if Table 1’s budgets are selected on an independent validation split. If tuned on test examples, they show an observed bound, not validated "iso-quality" preservation. Additionally, the main results fix tree size at 60, ignoring Fig 3's evidence that EAGLE can prefer smaller trees, failing to establish superiority over independently tuned baselines.
  2. Routing inconsistencies affect evidential confidence. Sec 2.1 states Qwen3 skips expert weight renormalization, but the official config enables norm_topk_prob. Furthermore, Fig 7’s caption claims 95% routing-probability coverage (MBPP, 8 experts), while the plot shows ~65%. These contradictions in model semantics and motivating mechanisms must be resolved with actual configs and measurements.
Questions for authors
  1. Which exact Qwen3 checkpoint and routing normalization were used for AR, EAGLE-3, and MoE-Spec? Please specify the normalization in Equation (8) and clarify whether correcting the Section 2.1 description changes any results.
  2. Can you provide the values and aggregation definition for Figure 7, and reconcile the approximately 65% plotted MBPP coverage at B=8 with the stated 95%?
  3. Were the budgets in Table 1 selected on a disjoint validation set? Please clarify the selection protocol and provide unrounded, paired quality differences with uncertainty estimates for the principal iso-quality comparisons.
  4. Do the main speed advantages persist when EAGLE and MoE-Spec independently select their tree sizes under the same validation and timing protocol? A focused comparison on representative settings, with absolute tokens/s, would help establish the practical magnitude of the benefit.
3 comments

Our shared response was posted before this review and answers the questions on budget selection and tree size with new results. We encourage you to take a look and ask any follow-up questions you may have.

"Which exact Qwen3 checkpoint and routing normalization were used for AR, EAGLE-3, and MoE-Spec?"

Thank you for catching this. The sentence in Section 2.1 has the two model names swapped. OLMoE does not renormalize the top-k weights and Qwen3 does. Our code reads the norm_topk_prob setting from the configuration of each checkpoint (false for OLMoE-1B-7B and true for Qwen3-30B-A3B in Table 1 and Qwen3-30B-A3B-Instruct-2507 in the shared response). AR, EAGLE-3 and MoE-Spec therefore all renormalize on Qwen3 and the correction changes no result. In Equation (8) each Qwen3 token divides the router probabilities of its k experts in the shortlist by their sum. We will fix the sentence and add this definition in the final version.

"Can you provide the values and aggregation definition for Figure 7, and reconcile the approximately 65% plotted MBPP coverage at B=8 with the stated 95%?"

As we replied to Reviewer cjiG, the bars of Figure 7 are correct and its caption and text were wrong. Coverage at budget B is the share of routing probability that the B experts with the most probability capture when we sum probability over the tokens of a draft tree. We average it over the 63-token draft trees of 50 prompts and over the 16 MoE layers of OLMoE-1B-7B. On MBPP the top 8 of 64 experts capture 64% of the routing probability and the top 24 capture 92%. Appendix D of the revision replaces the bars with a table of these values for all five benchmarks.

"Were the budgets in Table 1 selected on a disjoint validation set?"

No. We read the operating points of Table 1 off the test results. The shared response fixes the budget on training problems before any test run. Table G1 gives the paired accuracy differences on the test sets with 95% bootstrap intervals. At the conservative budget no task shows a significant accuracy change and the aggressive budget loses at most 1 point.

"Do the main speed advantages persist when EAGLE and MoE-Spec independently select their tree sizes under the same validation and timing protocol?"

Yes on the reasoning and code tasks. We tuned EAGLE-3 over 12 trees of up to 64 tokens and it runs fastest on a 64-token tree. Because MoE-Spec on its own fastest tree is at least as fast as on this tree, Table G1 gives a lower bound on its gain over the tuned EAGLE-3 on these tasks. We also timed MoE-Spec on trees of 96 and 128 tokens and only MATH500 ran faster there.

Table Q1 Throughput in tokens per second at batch 1. EAGLE-3 and MoE-Spec use the 64-token tree.

MethodGSM8KMATH500HumanEvalMBPP
Autoregressive116.8116.1115.4115.3
EAGLE-3175.1185.8175.8180.1
MoE-Spec (44 experts per layer)187.3198.6185.5193.1

Thanks for the response. I accept both corrections (the Section 2.1 OLMoE/Qwen3 swap and the Figure 7 caption error) and appreciate the disclosure about Table 1 budgets being read off test results. However, the new Table G1/G2 iso-quality gains of only 1.01–1.07× (0.99–1.04× at batch 16) are far from the 2.1→2.5× in Table 1, which does not support the paper's central claim. I will keep my original rating.

Thank you for reading our response. The comparison in your comment mixes two baselines. The 2.1x and 2.5x in Table 1 of the submission are speedups over autoregressive decoding. Over EAGLE-3 (the baseline of Table G1), Table 1 of the submission gives 1.19x on MATH500 and 1.08x and 1.10x on GSM8K and MBPP against 1.07x in SGLang. As your rating rests on this comparison, we ask that you revisit it on the same baseline or tell us which concern remains.

Comments on the paper

1 comment

We thank the reviewers for their careful reading. To answer the questions about our evaluation, we measured MoE-Spec in a production engine (SGLang with CUDA graphs on one A100) on the updated Instruct-2507 release of Qwen3-30B-A3B and the five benchmarks of the submission. Every method runs on the fastest EAGLE-3 tree that we found by tuning in SGLang. We fix the conservative budget in advance as the cheapest one that stays within 1 point of EAGLE-3 accuracy on GSM8K and MBPP training problems. A smaller aggressive budget trades some accuracy for more speed. Appendix F of the revised PDF (changes in blue) collects these results for review and we will integrate them into the main paper in the final version.

Table G1 Batch 1 at greedy decoding (T = 0) on the fastest EAGLE-3 tree. The conservative budget keeps 44 experts per MoE layer in each verification pass and the aggressive budget keeps 32. Each cell gives the speedup over EAGLE-3, then the accuracy change in points with its 95% interval. CNN/DM reports ROUGE-L.

TaskConservative (44)Aggressive (32)
GSM8K1.07x, -0.4 [-0.9, +0.1]1.16x, -1.0 [-1.7, -0.4]
MATH5001.07x, -0.5 [-2.0, +0.9]1.20x, -0.7 [-2.6, +1.1]
HumanEval1.06x, +0.4 [-0.6, +1.6]1.19x, +0.0 [-2.8, +3.0]
MBPP1.07x, +0.3 [-1.5, +2.3]1.20x, -0.7 [-3.3, +1.7]
CNN/DM1.01x, -0.1 [-0.3, +0.1]1.09x, +0.0 [-0.2, +0.1]

These gains are smaller than the Hugging Face gains in Table 1 of the submission. The EAGLE-3 paper reports the same gap for its own speedups and notes that efficiency gains are harder to obtain in highly optimized production engines such as SGLang (Li et al. 2025).

Table G2 Speedup over EAGLE-3 at each batch size with the budget in experts per layer. Parentheses give the accuracy change in points. Both methods run on the fastest EAGLE-3 tree for each batch size and this tree gets smaller as the batch grows.

BatchBudgetGSM8KMATH500HumanEvalMBPPCNN/DM
1Conservative (44)1.07x (-0.4)1.07x (-0.5)1.06x (+0.4)1.07x (+0.3)1.01x (-0.1)
Aggressive (32)1.16x (-1.0)1.20x (-0.7)1.19x (+0.0)1.20x (-0.7)1.09x (+0.0)
2Conservative (44)1.07x (-0.4)1.07x (-0.5)1.06x (+0.4)1.06x (+0.3)1.03x (-0.1)
Aggressive (32)1.14x (-1.0)1.12x (-0.7)1.14x (+0.0)1.14x (-0.7)1.06x (+0.0)
4Conservative (64)1.09x (-0.5)1.10x (+0.1)1.08x (-0.2)1.09x (-0.3)1.01x (+0.0)
Aggressive (48)1.19x (-0.7)1.14x (+0.6)1.17x (-0.8)1.14x (-2.6)1.04x (+0.0)
8Conservative (72)1.09x (-0.5)1.10x (-0.6)1.10x (+0.4)1.08x (-0.1)1.02x (+0.0)
Aggressive (64)1.13x (-0.8)1.13x (+0.4)1.13x (+0.6)1.13x (-2.7)1.03x (+0.0)
16Conservative (96)1.04x (-0.4)1.00x (-0.2)1.00x (+0.4)1.02x (+0.5)0.99x (+0.0)
Aggressive (80)1.05x (-0.7)1.05x (-0.1)1.09x (+0.4)1.07x (-0.8)1.01x (+0.0)

The gain shrinks as the batch grows because each request brings its own draft tree and different requests route to different experts. At batch 8 the budget still removes about a fifth of the experts that a pass would load. At batch 16 the budget must keep nearly all of them to stay within 1 point and little gain remains.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.