acceptodds
Under review as a conference paper at ICLR 2027

Composable Decoding on the Probability Simplex: Theory and Implementation

Abstract

Decoding for large language models is typically treated as a collection of isolated sampling strategies, with limited theoretical understanding of the behaviours they induce and how their underlying objectives relate. We formulate decoding as an optimisation problem over next-token distributions on the probability simplex, balancing expected model score against regularisation under support constraints. This view recovers familiar decoding methods through choices of regularisers and support constraints; more importantly, it enables new decoders to be constructed by composing distributional preferences within a single optimisation problem without external rewards, learned critics, or model parameter updates. We introduce CompoSimplex, a library with configurable support rules, regularisation primitives, and simplex solvers for constructing and evaluating compositional decoders. We evaluate standard samplers, individual regularisers, and compositions across multiple models and reasoning tasks. Our results show that compositions can realise trade-offs between single-sample quality, multi-sample quality, and diversity that are not attained by individual decoding objectives.

Primary area:
datasets and benchmarks
Keywords:
decoding, large language models, regularised optimisation

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_4B6DOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces a unified mathematical framework and library for "composable" decoding. The authors formulate decoding as an optimization problem with a flexible, general regularization term that allows one to bias the output distribution in different ways. They solve this optimization problem on the simplex with mirror ascent.

They show that several well-known existing decoding methods (such as temperature, top-p, top-k, etc.) can be seen as special cases of this regularization term. However, the flexibility of the formulation allows one to further compose multiple objectives together with weights. To explore this, the authors then introduce a new decoding method, "Best-of-KK Decoding", which jointly optimizes for consistency and coverage.

The paper evaluates the proposed framework and decoding method in 4 diverse model-benchmark settings, and they show modest-to-large gains across GPQA, Math-500, IFEval, and LiveCodeBench, compared to vanilla single-objective decoding methods such as top-p, top-k, etc. Most notably, jointly optimizing Jenson-Shannon divergence and diversity allows them to achieve a +10.6 pp gain on GPQA pass@1 for Qwen3-4B-Base.

Critical strengths
  1. The core framework is simple, clean, and quite general. This makes the space of decoding methods much easier to reason about, and the formulation is intuitive once stated.

  2. The compositional interface is flexible and naturally extensible. New differentiable preferences can be introduced as regularizers, combined with existing ones, weighted, and optimized through the same general machinery, rather than requiring a new decoding algorithm for each setting. Although the particular regularizers explored in this paper are fairly simple, the framework seems broad enough to support substantially richer objectives in future work. For example, one could imagine objectives for open-ended writing or other settings with several competing desiderata that need to be balanced jointly. More broadly, I see value in having a common framework in which new decoding objectives can be incorporated and composed without changing the overall decoding machinery.

  3. The empirical results show that the framework can find meaningfully better operating points. The strongest pass@1 improvements are substantial (e.g., +10.6 points on GPQA with Top-k support and +5.4 points on MATH500 with Top-k), and other settings show intuitive trade-offs between single-sample quality, multi-sample success, and diversity.

Critical weaknesses
  1. The empirical scope is a little limited (without trying to be "Reviewer 2"). The paper evaluates four model/benchmark pairs covering math, scientific reasoning, instruction following, and code, but does not test open-ended generation or writing quality. I would have liked to see at least one setting where the objective trades off more subjective generation properties, since this seems like a particularly natural use case for the framework.
  2. The gains from composition itself are not uniformly large. In several settings, a single regularizer is enough to realize most of the improvement over the base sampler, and the composed objective mainly provides a better trade-off between metrics rather than a large absolute gain.
  3. One other minor concern: objectives without closed-form solutions require iterative mirror ascent and add some inference overhead.
Questions for authors

Thank you for submitting an interesting paper!

One clarifying question: Most of the main experiments use fixed decoding hyperparameters (e.g., T=0.5T=0.5, λ=1\lambda=1, and equal composition weights). How much of the reported improvement, particularly the pass@1 gains, is still left over when the standard samplers and single-regularizer baselines are individually tuned? I think this would help clarify how much of the gain comes specifically from composition rather than from moving to a better point in the decoding hyperparameter space.

On presentation: I found the main results in Figure 2 somewhat difficult to parse at a glance, especially when comparing the base sampler, individual regularizers, and compositions across settings. The plots are useful for showing trade-offs, but make it harder to see the magnitude of the gains or which method is strongest on a particular metric. To improve presentation and clarity, perhaps the authors could include a compact summary table in the main body (for example, comparing the best standard sampler, single regularizer, and composition for each setting)? This would make the overall empirical picture much easier to follow.

3 comments

We thank the reviewer for the positive assessment and constructive suggestions. We will include additional evaluations of open-ended generation and tuned baselines in the revised manuscript.

The paper evaluates four model/benchmark pairs covering math, scientific reasoning, instruction following, and code, but does not test open-ended generation or writing quality. I would have liked to see at least one setting where the objective trades off more subjective generation properties, since this seems like a particularly natural use case for the framework.

We additionally evaluate Qwen2.5-7B on the poetry couplet completion task from FUDGE [1] to extend our evaluation to open-ended generation. Given the first line of the closing couplet of each of 154 Shakespearean sonnets, the model generates the next line. We report pass@1, pass@4, and pass@16 based on the benchmark's criteria, including iambic pentameter and rhyme. As shown in Table 1, our compositional sampler also shows improvements on this task with both base samplers. With Min-p sampling, KL+Diversity improves pass@1 from 9.7 to 16.9 while maintaining pass@16 at 63.0, while KL+Coverage improves pass@16 from 63.0 to 66.2. With Top-p sampling, KL+Coverage improves pass@1 from 12.3 to 16.2 and pass@16 from 65.6 to 68.2. We will include these additional results on constrained open-ended generation in the revised manuscript.

TaskMethodPass@1Pass@4Pass@16
FUDGE PoetryMin-p9.732.563.0
FUDGE PoetryMin-p KL12.333.159.1
FUDGE PoetryMin-p Diversity14.936.463.6
FUDGE PoetryMin-p JS11.733.163.0
FUDGE PoetryMin-p Coverage9.729.962.3
FUDGE PoetryMin-p KL + Diversity16.935.763.0
FUDGE PoetryMin-p KL + Coverage12.337.066.2
FUDGE PoetryMin-p JS + Coverage13.635.163.0
FUDGE PoetryTop-p12.327.365.6
FUDGE PoetryTop-p KL15.636.462.3
FUDGE PoetryTop-p Diversity14.934.464.9
FUDGE PoetryTop-p JS14.935.766.2
FUDGE PoetryTop-p Coverage14.337.066.9
FUDGE PoetryTop-p KL + Diversity16.236.466.2
FUDGE PoetryTop-p KL + Coverage16.236.468.2
FUDGE PoetryTop-p JS + Coverage16.235.766.9

Table 1: Evaluation on open-ended generation using Qwen2.5-7B

[1] Yang, Kevin, and Dan Klein. "FUDGE: Controlled text generation with future discriminators." Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021

In several settings, a single regularizer is enough to realize most of the improvement over the base sampler, and the composed objective mainly provides a better trade-off between metrics rather than a large absolute gain.

How much of the reported improvement, particularly the pass@1 gains, is still left over when the standard samplers and single-regularizer baselines are individually tuned?

We agree that a single regularizer can account for much of the improvement in a particular metric, while different regularizers can provide gains on different metrics. Composition can combine these complementary benefits and achieve a better balance across different metrics using a single configuration.

We tuned the hyperparameters of standard samplers, individual regularizers, and their composition on a small validation set using Optuna, and evaluated the selected configurations on the MATH500 test set. As shown in Table 2, our composition achieves 65.2 pass@1, compared with 59.8 for tuned Top-k and 64.8 for KL, the best-performing tuned single regularizer. Although the pass@1 gain over KL is relatively small, our composition also improves pass@4 from 79.6 to 81.6 and pass@16 from 86.6 to 89.4, showing a better balance across different sampling budgets. We also report additional temperature-tuning results across four benchmarks in Table 3. Across all eight configurations, our composable sampler, without task-specific temperature tuning, improves pass@16 over the pass@1-tuned baseline and pass@1 over the pass@16-tuned baseline. For example, on MATH500 with min-p, the pass@1-tuned baseline achieves (64.4, 84.4) and the pass@16-tuned baseline achieves (46.2, 89.6), whereas our composition achieves (67.2, 90.2) in a single configuration. We will include these additional results and tuning details in the revised manuscript and release the hyperparameter tuning interface.

MethodConfigurationPass@1Pass@4Pass@16
Top-kk=200k=20059.881.290.0
KLk=500,λ=0.27k=500, \lambda=0.2764.879.686.6
JSk=50,λ=0.17k=50, \lambda=0.1763.480.689.0
Coveragek=300,λ=1.11k=300, \lambda=1.1162.281.688.8
Diversityk=100,λ=0.50k=100, \lambda=0.5062.481.689.2
Compositionk=200,λ=1.00,αJS=0.50,αCoverage=0.25,αDiversity=0.25k=200, \lambda=1.00, \alpha_{\text{JS}}=0.50, \alpha_{\text{Coverage}}=0.25, \alpha_{\text{Diversity}}=0.2565.281.689.4

Table 2: Comparison of tuned standard samplers, individual regularizers, and composition on MATH using Qwen2.5-7B

TaskMethodPass@1Pass@4Pass@16
MATH500Top-k tuned for pass@1 (T=0.10)63.077.286.0
MATH500Top-k tuned for pass@16 (T=0.25)62.882.092.0
MATH500Top-k KL + Diversity (T=0.25)65.284.092.0
MATH500Top-k KL + Diversity (T=0.50)64.481.490.0
MATH500Min-p tuned for pass@1 (T=0.10)64.476.284.4
MATH500Min-p tuned for pass@16 (T=1.00)46.277.889.6
MATH500Min-p JS + Coverage (T=0.50)67.284.690.2
GPQA DiamondTop-k tuned for pass@1 (T=0.25)24.748.575.3
GPQA DiamondTop-k tuned for pass@16 (T=1.00)14.149.586.9
GPQA DiamondTop-k KL + Diversity (T=0.50)21.250.583.8
GPQA DiamondTypical tuned for pass@1 (T=0.50)20.757.680.3
GPQA DiamondTypical tuned for pass@16 (T=1.00)17.243.480.8
GPQA DiamondTypical KL + Coverage + Diversity (T=0.50)25.360.183.3
IFEvalTop-p tuned for pass@1 (T=0.25)52.167.575.6
IFEvalTop-p tuned for pass@16 (T=1.00)50.372.384.3
IFEvalTop-p JS + Entropy (T=0.50)53.170.480.0
IFEvalη-sampling tuned for pass@1 (T=0.25)54.263.272.6
IFEvalη-sampling tuned for pass@16 (T=1.00)49.071.383.2
IFEvalη-sampling KL + Coverage (T=0.50)51.972.383.4
LiveCodeBench v6Top-p tuned for pass@1 (T=0.10)58.362.967.4
LiveCodeBench v6Top-p tuned for pass@16 (T=1.00)53.765.169.7
LiveCodeBench v6Top-p KL + Diversity (T=0.50)56.665.768.6
LiveCodeBench v6Min-p tuned for pass@1 (T=0.10)56.662.365.7
LiveCodeBench v6Min-p tuned for pass@16 (T=1.00)54.364.667.4
LiveCodeBench v6Min-p JS + Diversity (T=0.50)54.965.169.7

Table 3: Comparison of Composable Sampler with temperature-tuned baselines across four benchmarks.

One other minor concern: objectives without closed-form solutions require iterative mirror ascent and add some inference overhead.

We acknowledge that iterative mirror ascent introduces additional inference overhead, with detailed discussion in Section 5.1 and Appendix D.3. We note that an LLM forward pass has O(Ld2+Lsd+Vd)O(Ld^2+Lsd+Vd), where LL, dd, ss, and VV denote the number of layers, hidden dimension, context length, and vocabulary size. Each mirror-ascent update has O(n)O(n), where nn is the sampling support size, resulting in O(mn)O(mn) for mm iterations. In our main experiments, we use m=10m=10, with m≪dm\ll d and n≤Vn\leq V (e.g., n=200n=200 for the default Top-k configuration). Thus, for the regularizers considered in this paper, the additional optimization does not change the overall asymptotic complexity of decoding, does not directly depend on the number of layers LL and the hidden dimension dd, and requires no additional model forward passes. The number of iterations also provides a trade-off between computational cost and optimization accuracy, we will further clarify this in the revised manuscript.

I found the main results in Figure 2 somewhat difficult to parse at a glance, especially when comparing the base sampler, individual regularizers, and compositions across settings. The plots are useful for showing trade-offs, but make it harder to see the magnitude of the gains or which method is strongest on a particular metric. To improve presentation and clarity, perhaps the authors could include a compact summary table in the main body? This would make the overall empirical picture much easier to follow.

We agree that a summary table would make it easier to compare the performance of different methods. Figure 2 is mainly intended to illustrate the trade-offs between different sampling methods across tasks. We will include a compact table with simplified results in the main text of the revised manuscript to improve readability and highlight the magnitude of the improvements. We thank the reviewer for this suggestion.

Thank you so much for the new results regarding open-ended generation, baseline tuning, and presentation! This addresses most of my major concerns, and I'm happy to maintain my strong accept recommendation.

Reviewer_4Dv9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper develops a framework for language-model decoding based on regularized optimization over the probability simplex. The authors formulate next-token decoding as maximizing expected model scores subject to distributional regularizers and support constraints. They show how standard methods, including softmax, Top-k, Top-p, and sparsemax, can be represented within this formulation.

The main proposed extension is compositional decoding, where multiple regularization terms are optimized jointly. The authors derive first-order optimality conditions and employ mirror ascent for objectives without convenient closed-form solutions. They also introduce Best-of-K (BoK) decoding, combining KL regularization with a weighted token-coverage objective intended to improve diversity across multiple samples.

The CompoSimplex library implements these components and integrates them with language-model generation backends. Experiments across MATH500, GPQA Diamond, IFEval, and LiveCodeBench compare individual regularizers and their compositions using single-sample accuracy, multi-sample success, and diversity metrics. While several compositions achieve favorable empirical trade-offs, the theoretical novelty and the extent to which these improvements are attributable specifically to composition remain uncertain.

Critical strengths
  1. Coherent mathematical formulation and flexible implementation. The paper presents a systematic way to express decoding through constrained regularized optimization. The recovery of familiar decoding distributions and the use of mirror ascent provide a consistent mathematical foundation. The BoK coverage objective is well defined for independent token draws at a fixed prefix, and its diminishing-returns structure is mathematically meaningful. The accompanying library makes the proposed formulations practically accessible without model retraining or additional forward passes.

  2. Broad empirical evaluation and useful diagnostic analyses. The experiments examine several models, benchmarks, support rules, and decoding objectives. Some compositions achieve meaningful trade-offs between accuracy and diversity. For example, KL+Diversity on MATH500 preserves KL's 64.4% pass@1 while improving pass@16 from 88.4% to 90.0%. The inclusion of temperature sweeps, coefficient sensitivity analyses, repeated-seed measurements, numerical optimization diagnostics, and inference-cost measurements strengthens the experimental presentation.

Critical weaknesses
  1. ** insufficient mathematical justification of the proposed compositional objective.**

The optimization formulation in Eq. (1), the weighted regularizer construction in Eqs. (5)–(7), and the multiplicative mirror-ascent update in Eq. (11) are largely direct applications of established regularized prediction and mirror-descent principles. Although the derivations appear correct, the paper does not establish a substantial new theoretical result explaining when or why composition provides capabilities unavailable to simpler decoding objectives.

More specifically, some proposed compositions are mathematically redundant. Suppose the reference distribution satisfies log p(v) = κs(v) + c, as in Appendix C.2 when the reference and scores are derived from the same logits. For a weighted KL and negative-entropy objective with coefficients a and b, respectively, the optimizer satisfies

q*(v) ∝ exp(((1/λ + aκ)/(a+b))s(v)).

Thus, combining KL and entropy produces exactly a temperature-scaled softmax distribution with an effective temperature. Under the same assumption, adding Diversity to this combination is equivalent to an Entropy+Diversity objective with reparameterized coefficients. This raises an important expressivity question: which compositions genuinely enlarge the decoder family, rather than merely change existing hyperparameters? No systematic equivalence characterization is provided.

A second mathematical limitation concerns Eq. (12). Its coverage utility correctly represents the probability of observing token v among K independent samples at a fixed prefix. However, during autoregressive generation, the K sampled trajectories generally diverge, and the number of trajectories reaching a later prefix is not necessarily K. The relevant occupancy probability at such a prefix depends on its actual visit count N, giving 1−(1−q(v))^N rather than the expression using a fixed K.

More fundamentally, maximizing local token coverage need not improve complete-sequence best-of-K success. Consider two alternatives, a and b, where only a produces a correct completion. For K=2, q(a)=0.9 gives success probability 0.99, whereas q(a)=0.5 gives success probability 0.75. Nevertheless, equal-weight token coverage is higher under the uniform distribution. This counterexample does not invalidate the proposed local surrogate, but demonstrates that its optimization is not inherently aligned with the intended downstream metric. The paper provides no bound, sufficient condition, or approximation analysis establishing such alignment.

Finally, the general claims surrounding KKT conditions and mirror ascent require stronger assumptions. For arbitrary user-defined regularizers, first-order stationarity need not characterize a global optimum, and convergence of the iterative solver is not automatic. Convexity, smoothness, and step-size assumptions should be specified for the claimed optimization properties. The absence of a new mathematical guarantee, together with the surrogate mismatch, substantially weakens the theoretical contribution.

  1. The experimental comparisons do not isolate composition's contribution, and the reported optimization accuracy is insufficient to establish objective-level conclusions.

In Eq. (6), an equally weighted two-component composition at strength λ assigns coefficient λ/2 to each constituent, while an individual primitive at the same λ receives the full coefficient λ. Therefore, comparisons at equal global strength are not comparisons at equal effective utility strength.

This directly affects the central robustness interpretation in Section 5.2. In Table 8, pure Coverage at λ=2 obtains 41.0% pass@1, compared with 60.2% for KL+Coverage and 62.6% for JS+Coverage. However, pure Coverage at λ=1 already obtains 63.4% pass@1, exceeding both compositions at λ=2 while using the same effective coverage coefficient. Thus, the reported advantage under strong regularization cannot be attributed solely to the addition of a divergence term. Appropriately coefficient-matched comparisons are essential.

The numerical implementation introduces a further concern. Appendix D.3 reports mean L1 distances of approximately 0.033–0.062 from reference optima for the 10-step mirror-ascent configuration used in the task experiments. These are non-negligible local discrepancies. More importantly, convergence in average next-token L1 distance on 128 cached prefixes does not establish fidelity of the resulting autoregressive sequence distribution. Conditional errors can accumulate across generation steps, and the reported average does not provide a uniform bound over encountered trajectories. The experimental observations may therefore reflect both the intended regularization objective and the particular finite-step optimization dynamics.

There is also a mathematical comparability issue in the distributional analysis. Appendix D.1 averages KL, JS, coverage, and diversity metrics along trajectories generated by each decoder. Different decoders generally visit different prefixes, so the reported quantities combine changes in the next-token distribution with changes in the distribution over visited contexts. Consequently, Figure 3 does not isolate pointwise regularization effects on a common set of prefixes, and its empirical Pareto interpretation is limited.

many performance differences are small, statistical replication is concentrated on MATH500, and iterative decoders incur approximately 1.13–2.88× inference overhead. Without stronger tuned baselines, uncertainty estimates, near-converged solver comparisons, and compute-matched evaluation, the practical advantage of composition remains insufficiently established.

Finally, I should mention that there other more entropy/geometry based decoding methods that you didn't include in your experiments. There are geometry based decodings, and entropy based decodings.

Questions for authors

very important — 1. Mathematical expressivity and novelty

Given that KL+Entropy reduces exactly to temperature scaling when the reference log probabilities are affine in the model scores, which proposed compositions are mathematically different from simpler decoder families? Can the authors clarify under what conditions composition produces genuinely new optimal distributions rather than just reparameterizing existing objectives?

important — 2. Relationship between local coverage and sequence-level success

Equation (12) assumes K independent draws at a shared prefix, but autoregressive trajectories diverge, and later prefixes may be visited by fewer than K samples. What is the theoretical justification for using the same K at every generation step? Can the authors provide any conditions or bounds connecting improvements in local token coverage to actual sequence-level pass@K, given that increasing token coverage can sometimes reduce success probability?

Important — 3. Coefficient-matched experimental comparisons

How would the main results change if individual primitives and compositions were compared using the same effective utility coefficients instead of the same global λ? In Table 8, Coverage at λ=1 achieves 63.4% pass@1, outperforming KL+Coverage and JS+Coverage at λ=2 despite having the same coverage coefficient. How do the authors explain the claimed robustness advantage in this case? Are there consistent improvements beyond temperature tuning and adjusting regularization strength?

Important — 4. Optimization accuracy and comparability

Do the reported improvements remain when the iterative objectives are solved more accurately? Given the nonzero distributional error of the 10-step solver, how sensitive are the sequence-level results to optimization accuracy? Also, would the distributional comparisons change if all decoders were evaluated on the same cached prefixes rather than their own generated trajectories? Do the conclusions hold under matched computational budgets?

  1. Convexity, uniqueness, and global optimality

What assumptions on the regularizers guarantee existence, uniqueness, and global optimality in Eq. (6)? The KKT conditions in Eqs. (3)–(7) are generally necessary but not sufficient for nonconvex objectives. Since the framework allows arbitrary user-defined regularizers, can the authors clarify which guarantees hold for the implemented primitives and which require additional convexity, smoothness, or boundary assumptions?

  1. Mirror-ascent convergence and numerical stability

What convergence guarantees apply to Eq. (11) for the proposed combinations of KL, JS, and coverage regularizers? How is the fixed step size justified when the objective's curvature changes with λ, K, and composition weights? Also, since multiplicative updates cannot recover a coordinate once its probability becomes exactly zero, how does the implementation handle numerical underflow and prevent valid tokens from being permanently excluded?

  1. Scaling and comparability of regularization terms

The regularization terms in Eq. (6) have different scales, gradient magnitudes, and sensitivities to support size and K. How are their coefficients normalized so that equal composition weights represent comparable regularization strengths? In particular, the coverage gradient contains K(1−q(v))^(K−1), which changes with K. Could some of the observed improvements come from differences in gradient magnitudes rather than the benefits of composition itself?

  1. Dependence of BoK on the sampling budget

How sensitive is the proposed BoK objective to the choice of K? When K=1, the coverage utility becomes linear in q and can be absorbed into a modified token-score vector. What changes mathematically when K>1, and is there any theoretical or empirical advantage as the sampling budget increases? How should K be chosen when the actual inference budget differs from the configured value?

  1. Difference between optimized and evaluated coverage

Appendix D.1 reports coverage using a normalized utility over the top eight reference tokens, while the optimization objective uses a potentially different candidate set and weighting scheme. How closely related are these two definitions? Can they rank decoding distributions differently? Would the reported improvements remain if exactly the same coverage definition were used for optimization and evaluation?

  1. Validity of the Pareto interpretation

Figure 3 presents several compositions as non-dominated in divergence–utility space. However, Appendix D.1 averages these metrics over prefixes visited by each decoder, meaning the compared distributions are evaluated on different contexts. Would the claimed trade-offs remain if all decoders were evaluated on the same fixed set of prefixes? How is Pareto dominance defined when the measured objectives also depend on which trajectories the decoder generates?

  1. Exact numerical solutions versus approximate decoding

Appendix D.3 computes reference optima using the KKT conditions and numerical bisection. For the implemented convex objectives, why not use these more accurate solutions during generation? Is there a clear computational advantage to fixed-step mirror ascent over a specialized numerical solver at similar accuracy? A comparison of objective error, distributional error, and runtime would help justify the choice of solver.

  1. Statistical significance and selection effects

How many decoder configurations and hyperparameter settings were tested before selecting the representative compositions in Figure 2? Several reported improvements are only one or two percentage points, and GPQA Diamond and LiveCodeBench have relatively small evaluation sets. Which improvements remain statistically significant under paired confidence intervals and multiple-comparison corrections? How can the authors distinguish consistent improvements from sampling variation or configuration selection?

very important — 13. Other decoding methods

Why are the comparisons limited to Top-k, Top-p, Min-p, typical and η-sampling? There are other recent decoding methods, especially entropy-aware approache like Top-H (Potraghloo et al., NeurIPS 2025) and geometry aware approach Top-W (Davoodi et al., ICML 2026). Given the paper's focus on diversity, entropy, and decoding quality, why are these relevant and more recent baselines missing? Would the proposed compositions still show improvements over Top-H and Top-W with comparable tuning and computational cost?

5 comments

We thank the reviewer for the detailed questions and suggestions. We first address the questions marked as "(very) important", followed by the remaining questions. We will incorporate the additional theoretical discussion and experimental results into the revised manuscript.

Given that KL+Entropy reduces exactly to temperature scaling when the reference log probabilities are affine in the model scores, which proposed compositions are mathematically different from simpler decoder families? Can the authors clarify under what conditions composition produces genuinely new optimal distributions rather than just reparameterizing existing objectives?

We first give a criterion for when a composition produces new optimal distributions, and then apply it to our regularisers to say which compositions differ from simpler families.

Proposition R1. Let q⋆q^\star be the optimum of a composition of convex, differentiable regularisers Ω1,…,Ωm\Omega_1,\dots,\Omega_m with positive coefficients θi=λαi\theta_i=\lambda\alpha_i, and suppose q⋆(v)>0q^\star(v)\gt 0 for every supported token (this holds when Entropy, KL or JS is present). Let 1\mathbf{1} be the all-ones vector.

(i) If the gradients ∇Ω1(q⋆),…,∇Ωm(q⋆)\nabla\Omega_1(q^\star),\dots,\nabla\Omega_m(q^\star) and 1\mathbf{1} are linearly independent, then no composition of a strict subset of these regularisers has q⋆q^\star as an optimum, whatever its coefficients.

(ii) If instead one of the gradients is a linear combination of the others and 1\mathbf{1}, then q⋆q^\star satisfies the optimality condition (Eq. 7) of the composition without that regulariser, with modified coefficients, and it is optimal for that composition when these coefficients are nonnegative.

Both parts follow from Eq. (7), which is linear in the coefficients: (i) is the uniqueness of KKT multipliers under linear independence [1, Sec. 12.3, Ex. 12.17], and (ii) is the sufficiency of the KKT conditions for convex problems [2, Sec. 5.5.3].

A regulariser cannot be dropped when its gradient at the optimum is not a combination of the gradients of the other terms and a constant, and it can be dropped, under the sign condition of (ii), when it is. The gradients are evaluated at q⋆q^\star, which Eq. (7) ties to the scores, so the test applies at the optimum and not to the formulas of the gradients. Importantly, it needs no closed form for q⋆q^\star.

For KL and Entropy, as Appendix C.2 notes, KL(q∥p)=−H(q)−⟨q,log⁡p⟩\mathrm{KL}(q\Vert p)=-H(q)-\langle q,\log p\rangle and log⁡p\log p is a positive affine transformation of the scores, so a KL term is an Entropy term at another temperature (case (ii) when both are present).

However, JS, Coverage and Diversity cannot be dropped. Let us count Entropy and KL as one term, with gradient log⁡q\log q. With a second term, the linear independence in (i) means that its gradient at the optimum is not of the form alog⁡q⋆+ca\log q^\star+c for constants aa and cc (and that q⋆q^\star is not uniform). This form is impossible for each of our three regularisers:

  • Coverage has gradient −Kw(1−q)K−1-Kw(1-q)^{K-1}, with weights ww that are zero outside the top-rr reference tokens. The gradient vanishes on those tokens and not on the others, which alog⁡q⋆+ca\log q^\star+c cannot do once two of them have different probabilities under pp.
  • Diversity has the same gradient, with weights that are zero on the most likely token and largest at intermediate gaps. It can take that form only in a degenerate case, when the tokens other than the most likely one share at most two probability values under pp.
  • JS has gradient 12log⁡2qq+p\tfrac12\log\frac{2q}{q+p}. The form is impossible once three tokens have different probabilities under pp, with one exception: when the Entropy coefficient is Tref/TT_{\mathrm{ref}}/T where TT is the temperature of the scores and TrefT_{\mathrm{ref}} that of the reference pp, we get q⋆=pq^\star=p and the gradient vanishes.

Hence KL + Coverage, KL + Diversity and JS + Entropy are neither temperature scaling nor the JS or Coverage or Diversity term alone. For compositions with two of these terms (JS with Diversity, JS with Coverage), we tested Proposition R1(i) on 128 fixed MATH500 prefixes with Qwen2.5-7B. We used KL + Entropy as a control because it is equivalent to temperature scaling when the reference log probabilities are affine in the model scores. On every tested prefix, both JS mixtures had linearly independent projected gradients. Their optima could not be reproduced by retuning either component alone or by changing the temperature on the same candidate set.

CompositionIndependentBoth requiredTemperature equivalent
KL + Entropy0/1280/128128/128
JS + Coverage128/128128/1280/128
JS + Diversity128/128128/1280/128

Table 1: Evaluation of gradient independence, regularizer non-redundancy, and temperature equivalence

[1] J. Nocedal and S. J. Wright, Numerical Optimization, 2nd ed., Springer, 2006. [2] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004.

Equation (12) assumes K independent draws at a shared prefix... Can the authors provide any conditions or bounds connecting improvements in local token coverage to actual sequence-level pass@K, given that increasing token coverage can sometimes reduce success probability?

How sensitive is the proposed BoK objective to the choice of K? ... How should K be chosen when the actual inference budget differs from the configured value?

We sample complete answers independently. In Equation (12), K is a reference sampling budget used to evaluate the next-token distribution at a given prefix, rather than the actual number of samples reaching that prefix. If only n samples reach a prefix, the probability of observing token v at least once is 1−(1−q(v))n1-(1-q(v))^n. Using a fixed local K provides a consistent coverage objective across decoding steps, but does not directly estimate sequence-level success.

The relationship between local coverage and sequence-level success depends on the quality of the subsequent continuations. For a fixed prefix hh, if only its next-token distribution changes from qq to q′q', the change in single-sequence success probability follows from the law of total probability: Δp=P(h)∑v[q′(v∣h)−q(v∣h)]ah(v)\Delta p =P(h)\sum_v [q'(v\mid h)-q(v\mid h)]a_h(v), where ah(v)a_h(v) is the probability of generating a correct completion after selecting token vv. Increasing local coverage shifts probability mass from the most likely tokens toward less likely alternatives. A sufficient condition for this shift to give Δp>0\Delta p>0 is that every token gaining mass has a higher continuation success probability than every token losing mass: ah(v)>ah(u)a_h(v)>a_h(u) whenever q′(v∣h)>q(v∣h)q'(v\mid h)>q(v\mid h) and q′(u∣h)<q(u∣h)q'(u\mid h)<q(u\mid h). Since complete generations are independent, a prompt with single-sequence success probability pp has pass@B=1−(1−p)B\mathrm{pass@B}=1-(1-p)^B, which is increasing in pp, so under this condition pass@B increases as well. Because aha_h is unknown at decoding time, higher local coverage does not in general guarantee better sequence-level success.

We further examined this relationship by varying the local K while keeping the number of complete generations fixed at 16. As shown in Table 2, increasing local K from 16 to 32 increases DivGap on both MATH500 and GPQA, while pass@16 decreases from 90.0 to 88.8 and from 83.84 to 83.33. We use the actual number of completions as a default choice for local K, while allowing it to be tuned on a validation set when optimizing for a different inference budget.

Local KKMATH500 DivGapMATH500 Pass@1/@4/@16GPQA DivGapGPQA Pass@1/@4/@16
80.078963.4/80.6/88.80.0485924.75/51.01/82.32
160.090964.4/81.4/90.00.0558521.21/50.51/83.84
320.104561.60/81.0/88.80.0692923.74/52.53/83.33

Table 2: Sensitivity to the local sampling budget KK

How would the main results change if individual primitives and compositions were compared using the same effective utility coefficients ... Are there consistent improvements beyond temperature tuning and adjusting regularization strength?

The regularization terms in Eq. (6) have different scales ... How are their coefficients normalized so that equal composition weights represent comparable regularization strengths?

We agree that equal composition weights can yield different gradient contributions. The gradient of Eq. (6) is ∇qf(q)=s−λ∑iαi∇qΩi(q),∑iαi=1\nabla_q f(q)=s-\lambda\sum_i\alpha_i\nabla_q\Omega_i(q), \sum_i\alpha_i=1. We normalize the composition weights to sum to one, so a global λ\lambda fixes the sum of the coefficients, while αi\alpha_i determines how that sum is allocated. However, equal weights do not necessarily imply equal effective regularization strengths. Even when one regularizer has the same effective coefficient, adding another regularizer changes the overall gradient and can lead to a different optimum. In our main comparisons, we fix the global λ\lambda to compare different allocations of the same total coefficient. We do not expect composition to outperform every individual regularizer on every metric. In Table 8 of the paper, Coverage alone achieves higher pass@1 (63.4) than KL+Coverage (60.2) and JS+Coverage (62.6) with matched effective coefficients. However, at global strength 2, Coverage alone drops to 41.0 pass@1, while the compositions achieve 60.2 and 62.6, illustrating better performance under this stronger regularization setting.

To better understand these coefficients, we additionally report the results of tuned standard samplers, individual regularizers, and their composition using Optuna. In Table 3, the selected KL and JS configurations use smaller λ\lambda values (0.27 and 0.17) than Coverage, Diversity and the compositional sampler (1.11, 0.5 and 1.0). In the composition, JS receives the largest weight, and this composition achieves 65.2% pass@1 and 89.4% pass@16, illustrating the trade-off between single-sample accuracy and success with a larger sampling budget.

We also report additional temperature-tuning results across four benchmarks in Table 4, showing that our compositional sampler retains improvements compared with temperature-tuned base samplers. Across all eight configurations, our composable sampler, without task-specific temperature tuning, improves pass@16 over the pass@1-tuned baseline and pass@1 over the pass@16-tuned baseline. For example, on MATH500 with min-p, the pass@1-tuned baseline achieves (64.4, 84.4) and the pass@16-tuned baseline achieves (46.2, 89.6), whereas our composition achieves (67.2, 90.2) in a single configuration. We will include these additional results and tuning details in the revised manuscript and release the Optuna-based hyperparameter tuning interface.

MethodConfigurationPass@1Pass@4Pass@16
Top-kk=200k=20059.881.290.0
KLk=500,λ=0.27k=500,\lambda=0.2764.879.686.6
JSk=50,λ=0.17k=50,\lambda=0.1763.480.689.0
Coveragek=300,λ=1.11k=300,\lambda=1.1162.281.688.8
Diversityk=100,λ=0.50k=100,\lambda=0.5062.481.689.2
Compositionk=200,λ=1.00,αJS=0.50,αCoverage=0.25,αDiversity=0.25k=200,\lambda=1.00,\alpha_{\text{JS}}=0.50,\alpha_{\text{Coverage}}=0.25,\alpha_{\text{Diversity}}=0.2565.281.689.4

Table 3: Comparison of tuned standard samplers, individual regularizers, and composition on MATH using Qwen2.5-7B

TaskMethodPass@1Pass@4Pass@16
MATH500Top-k tuned for pass@1 (T=0.10)63.077.286.0
MATH500Top-k tuned for pass@16 (T=0.25)62.882.092.0
MATH500Top-k KL + Diversity (T=0.25)65.284.092.0
MATH500Top-k KL + Diversity (T=0.50)64.481.490.0
MATH500Min-p tuned for pass@1 (T=0.10)64.476.284.4
MATH500Min-p tuned for pass@16 (T=1.00)46.277.889.6
MATH500Min-p JS + Coverage (T=0.50)67.284.690.2
GPQA DiamondTop-k tuned for pass@1 (T=0.25)24.748.575.3
GPQA DiamondTop-k tuned for pass@16 (T=1.00)14.149.586.9
GPQA DiamondTop-k KL + Diversity (T=0.50)21.250.583.8
GPQA DiamondTypical tuned for pass@1 (T=0.50)20.757.680.3
GPQA DiamondTypical tuned for pass@16 (T=1.00)17.243.480.8
GPQA DiamondTypical KL + Coverage + Diversity (T=0.50)25.360.183.3
IFEvalTop-p tuned for pass@1 (T=0.25)52.167.575.6
IFEvalTop-p tuned for pass@16 (T=1.00)50.372.384.3
IFEvalTop-p JS + Entropy (T=0.50)53.170.480.0
IFEvalη-sampling tuned for pass@1 (T=0.25)54.263.272.6
IFEvalη-sampling tuned for pass@16 (T=1.00)49.071.383.2
IFEvalη-sampling KL + Coverage (T=0.50)51.972.383.4
LiveCodeBench v6Top-p tuned for pass@1 (T=0.10)58.362.967.4
LiveCodeBench v6Top-p tuned for pass@16 (T=1.00)53.765.169.7
LiveCodeBench v6Top-p KL + Diversity (T=0.50)56.665.768.6
LiveCodeBench v6Min-p tuned for pass@1 (T=0.10)56.662.365.7
LiveCodeBench v6Min-p tuned for pass@16 (T=1.00)54.364.667.4
LiveCodeBench v6Min-p JS + Diversity (T=0.50)54.965.169.7

Table 4: Comparison of Composable Sampler with temperature-tuned baselines

There are other recent decoding methods, especially entropy-aware approache like Top-H (Potraghloo et al., NeurIPS 2025) and geometry aware approach Top-W (Davoodi et al., ICML 2026)

We additionally evaluated our compositional sampler with Top-H [3] using Qwen3-4B on GPQA Diamond. As shown in Table 5, composition achieves a similar improvement pattern to that observed with other base samplers. In particular, JS+Diversity improves pass@1 from 19.7 to 23.7 while maintaining pass@16 at 64.6. KL+Diversity improves pass@16 to 65.2. These results suggest that our framework can also combine complementary regularization objectives with recent sampling methods. We also attempted to evaluate Top-W [4], however, it required more inference time than Top-H in our current setup, making a complete evaluation difficult within the rebuttal period. We will include the additional Top-H results in the revised manuscript and discuss Top-W in the related work section.

MethodPass@1Pass@4Pass@16
Top-H19.745.564.6
Top-H + KL23.736.448.5
Top-H + JS23.239.454.0
Top-H + Diversity23.742.961.1
Top-H + KL + Diversity20.742.965.2
Top-H + JS + Diversity23.743.464.6

Table 5: Evaluation of compositional regularization with Top-H on GPQA Diamond using Qwen3-4B

[3] Baghaei Potraghloo, Erfan, et al. "Top-h decoding: Adapting the creativity and coherence with bounded entropy in text generation." Advances in Neural Information Processing Systems 38 (2026): 28482-28513.

[4] Davoodi, Arash Gholami, et al. "Geometry-Aware Decoding with Wasserstein-Regularized Truncation and Mass Penalties for Large Language Models." arXiv preprint arXiv:2602.10346 (2026).

Reviewer_3uycOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper focuses on decoding for large language models, specifically how to combine the behaviors of different sampling strategies within a single method. The authors frame each decoding step as an optimization problem over the next-token distribution, where the model’s score is balanced against a weighted sum of regularizers, such as KL divergence from the model distribution or a bonus for covering more tokens across samples. Standard samplers like Top-k and Top-p arise as special cases, and new decoders can be built by combining regularizers. The authors release CompoSimplex, a library implementing this framework for HuggingFace and vLLM. Experiments on four model-benchmark pairs covering math, science, instruction following, and code show that compositions can balance single-sample accuracy and multi-sample success better than their individual regularizers in some settings.

Critical strengths
  • The paper's CompoSimplex library provides a unified way to configure decoders under one interface, and it already integrates with two widely used generation libraries (Hugging Face Transformers and vLLM).
Critical weaknesses
  • All main comparisons use the base sampler at a single temperature (T=0.5), but Appendix C.2 shows that single regularizers like KL reduce exactly to temperature scaling, so some of the gains may come from an effectively different temperature rather than from composition. In Table 1, plain Top-k at T=0.25 reaches 92.0 pass@16 on MATH500, above every decoder evaluated at T=0.5. I would’ve liked to see a temperature-tuned baseline for each task.
  • Each benchmark is evaluated with a different model, so it's unclear whether differences across benchmarks reflect the task or the model. Evaluating at least one model across all four benchmarks would separate these effects. The paper also gives little guidance on choosing λ\lambda, the composition weights α\alpha, or the token weights. The sweep in Section 5.2 finds "no consistent improvement as α\alpha increases", which makes the framework hard to use in practice.
Questions for authors

I had no questions (apologies for copying in my review incorrectly, I wrote them separately and copied them into the wrong boxes)

3 comments

We thank the reviewer for the constructive feedback and helpful suggestions for improving this work.

All main comparisons use the base sampler at a single temperature (T=0.5) ... I would’ve liked to see a temperature-tuned baseline for each task.

We performed temperature tuning across four benchmarks and eight sampler configurations, jointly optimizing pass@1 and pass@16 on a validation set over T∈{0.1,0.25,0.5,0.7,1.0}T \in \{0.1, 0.25, 0.5, 0.7, 1.0\} using Optuna. We identified the empirical Pareto front and evaluated the configurations maximizing pass@1 and pass@16 on the test set. We observed a consistent temperature preference: 7/8 configurations favored lower temperatures T=0.1T=0.1 or T=0.25T=0.25 for pass@1, while 7/8 favored T=1.0T=1.0 for pass@16, consistent with the expected trade-off between single-sample accuracy and multi-sample success. Across all eight configurations, our composable sampler, without task-specific temperature tuning, improves pass@16 over the pass@1-tuned baseline and pass@1 over the pass@16-tuned baseline, as shown in Table 1. For example, on MATH500 with min-p, the pass@1-tuned baseline achieves (64.4, 84.4) and the pass@16-tuned baseline achieves (46.2, 89.6), whereas our composition achieves (67.2, 90.2) in a single configuration. These results demonstrate that composition can achieve a better joint performance than temperature scaling alone. We use T=0.5T=0.5 as a common reference in the main experiments to ensure consistent comparisons, and will include these comparisons in the revised manuscript.

Table 1: Comparison of Composable Sampler with temperature-tuned baselines

TaskMethodPass@1Pass@4Pass@16
MATH500Top-k tuned for pass@1 (T=0.10)63.077.286.0
MATH500Top-k tuned for pass@16 (T=0.25)62.882.092.0
MATH500Top-k + KL + Diversity (T=0.25)65.284.092.0
MATH500Top-k + KL + Diversity (T=0.50)64.481.490.0
MATH500Min-p tuned for pass@1 (T=0.10)64.476.284.4
MATH500Min-p tuned for pass@16 (T=1.00)46.277.889.6
MATH500Min-p + JS + Coverage (T=0.50)67.284.690.2
GPQA DiamondTop-k tuned for pass@1 (T=0.25)24.748.575.3
GPQA DiamondTop-k tuned for pass@16 (T=1.00)14.149.586.9
GPQA DiamondTop-k + KL + Diversity (T=0.50)21.250.583.8
GPQA DiamondTypical tuned for pass@1 (T=0.50)20.757.680.3
GPQA DiamondTypical tuned for pass@16 (T=1.00)17.243.480.8
GPQA DiamondTypical + KL + Coverage + Diversity (T=0.50)25.360.183.3
IFEvalTop-p tuned for pass@1 (T=0.25)52.167.575.6
IFEvalTop-p tuned for pass@16 (T=1.00)50.372.384.3
IFEvalTop-p + JS + Entropy (T=0.50)53.170.480.0
IFEvalη-sampling tuned for pass@1 (T=0.25)54.263.272.6
IFEvalη-sampling tuned for pass@16 (T=1.00)49.071.383.2
IFEvalη-sampling + KL + Coverage (T=0.50)51.972.383.4
LiveCodeBench v6Top-p tuned for pass@1 (T=0.10)58.362.967.4
LiveCodeBench v6Top-p tuned for pass@16 (T=1.00)53.765.169.7
LiveCodeBench v6Top-p + KL + Diversity (T=0.50)56.665.768.6
LiveCodeBench v6Min-p tuned for pass@1 (T=0.10)56.662.365.7
LiveCodeBench v6Min-p tuned for pass@16 (T=1.00)54.364.667.4
LiveCodeBench v6Min-p + JS + Diversity (T=0.50)54.965.169.7

Each benchmark is evaluated with a different model, so it's unclear whether differences across benchmarks reflect the task or the model. Evaluating at least one model across all four benchmarks would separate these effects.

We agree that evaluating the same model across different tasks can provide further evidence of the benefits of our approach, and ran additional experiments using Qwen2.5-7B on MATH500, GPQA Diamond, IFEval, and LiveCodeBench v6, using the same configuration as in the main experiments without task-specific tuning. As shown in Table 2, compared with the corresponding base samplers, our composition improves pass@1 and pass@4 on all tasks, while maintaining or improving pass@16. For example, on GPQA Diamond, pass@1 improves from 27.8 to 34.3 and pass@16 from 88.9 to 92.9; on IFEval, pass@4 improves from 47.1 to 51.9 and pass@16 from 60.3 to 65.1. These improvements are observed when the backbone model is fixed, indicating that the benefits of composition are not simply due to differences in the models used across benchmarks. The consistent improvements also suggest that the same configuration can be effective across different tasks without task-specific tuning.

Table 2: Cross-task evaluation of Qwen2.5-7B using shared hyperparameters

TaskMethodPass@1Pass@4Pass@16
MATH500Top-k59.881.290.0
MATH500KL64.480.688.4
MATH500Diversity56.079.491.0
MATH500KL+Diversity64.481.490.0
GPQA DiamondTop-k27.859.188.9
GPQA DiamondKL31.866.288.9
GPQA DiamondDiversity32.872.293.4
GPQA DiamondKL+Diversity34.371.292.9
IFEvalη-sampling30.747.160.3
IFEvalKL32.051.963.4
IFEvalDiversity32.551.865.1
IFEvalKL+Diversity32.051.965.1
LiveCodeBench v6Top-p14.321.726.3
LiveCodeBench v6KL16.621.726.9
LiveCodeBench v6Diversity14.920.627.4
LiveCodeBench v6KL+Diversity15.422.927.4

The paper also gives little guidance on choosing the composition weights or the token weights. The sweep in Section 5.2 finds "no consistent improvement as increases", which makes the framework hard to use in practice.

We would like to clarify that the effects of composition weights on sampling properties should be distinguished from their effects on accuracy. As shown in Figure 5, increasing the composition weight of regularizers monotonically improves the corresponding utility. This indicates that composition weights provide predictable control over sampling properties in practice. For example, increasing the weight of KL encourages the sampler to stay closer to the base distribution, while increasing the weight of Diversity places more emphasis on diversity. However, these choices do not necessarily lead to better pass@1 and pass@16 performance, even when the regularizer is related to the target evaluation metric in some way (e.g., Diversity and pass@16).

Our results (Figure 3 in Section 5.2 and Figure 5 in Appendix D.4) suggest that composition weights can be used to design compositional samplers with desired sampling properties. For samplers aiming to optimize specific metrics, such as pass@1 or pass@16, we will release an Optuna-based hyperparameter tuning interface that allows users to tune hyperparameters based on performance on a small validation set. This procedure applies to standard samplers as well as samplers with individual or compositional regularizers. As an example, Table 3 reports the results of hyperparameter tuning for Top-k-based samplers on MATH500, where regularization strength and other hyperparameters are selected based on a validation set. Compared with the tuned Top-k sampler, the tuned composition improves pass@1 from 59.8 to 65.2 while preserving pass@16 (89.4 vs. 90.0). The cross-task results in Table 2 further suggest that a shared configuration can achieve competitive performance without task-specific tuning. We will clarify these practical guidelines and report the selected hyperparameter configurations in the revised manuscript.

MethodConfigurationPass@1Pass@4Pass@16
Top-kk=200k=20059.881.290.0
KLk=500,λ=0.27k=500, \lambda=0.2764.879.686.6
JSk=50,λ=0.17k=50, \lambda=0.1763.480.689.0
Coveragek=300,λ=1.11k=300, \lambda=1.1162.281.688.8
Diversityk=100,λ=0.50k=100, \lambda=0.5062.481.689.2
Compositionk=200,λ=1.00,αJS=0.50,αCoverage=0.25,αDiversity=0.25k=200, \lambda=1.00, \alpha_{\text{JS}}=0.50, \alpha_{\text{Coverage}}=0.25, \alpha_{\text{Diversity}}=0.2565.281.689.4

Table 3: Comparison of tuned standard samplers, individual regularizers, and composition on MATH using Qwen2.5-7B

Thank you for the clarification on temperature scaling, as it was the main concern I generally have about sampling strategies. Based on these results, I will update my score to a weak accept.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.