PaperlayerView exact revision on arXiv

Source-linked AI summary

How code helps different tasks? A decompositional lens on LLM post-training

Zheng Yu, Yiwei Li, Yishen Chen, Xiang Li, Jiale Han, Benyou Wang, Jingbang Chen

arXiv:2609.33845v1cs.AI

TL;DR

The paper addresses the lack of evidence about which computational categories of code benefit particular models and tasks, and whether those benefits persist when categories are combined. It decomposes an execution-verified corpus and compares category-specific and compact-mixture fine-tuning. Compact mixtures outperform both their best constituent and full-corpus training on selected model–task pairs, while broader results vary by model and task.

  • Problem

    Existing findings about code exposure, representation, and complexity do not specify which computational categories benefit a given model or how their benefits change when trained together.

    Source: §1 · p. 2

  • Method

    The study groups execution-verified code solutions by computational pattern, compares single-category and balanced-mixture fine-tuning across models and tasks, then evaluates compact mixtures against constituents and full-corpus training.

    Source: §1 · p. 2, §3.1 · p. 3, §3.3 · p. 4

  • Results

    On selected model–task pairs, compact mixtures of individually beneficial categories outperform both their best constituent and full-corpus training while using roughly 10–15% of the full corpus.

    Source: §9 · p. 9

  • Takeaways & Limitations

    The value of code categories and their compositions depends on the starting model and target task, rather than broad category coverage alone ensuring stronger performance.

    Source: §1 · p. 2, §9 · p. 9

    The exploratory mixture screen uses the same evaluation suite for ranking and evaluation, selects qualifying cases after training, and does not establish optimal composition or causal synergy.

    Source: §8 · p. 9

Abstract

from arXiv · show

Evaluating code data as a single corpus can obscure which types of code data benefit which models and downstream tasks. Effective data selection requires understanding both the benefits of individual categories and whether these benefits persist when categories are combined. We introduce a decompositional lens for studying these effects in LLM post-training. We first decompose an execution-verified code corpus into interpretable categories based on the computational patterns of its solutions. Through controlled fine-tuning experiments, we compare individual categories with a balanced mixture across instruction-tuned models on question answering, mathematics, and code generation. The resulting response maps reveal recurring gains in average question-answering performance, while the same category can improve one model or task and degrade another. The best-performing category also varies with the starting model and target task. We then compose compact mixtures guided by these results and examine whether benefits observed in individual categories persist under joint training. On selected model--task pairs, mixtures whose constituents each improve the target task outperform both their best constituent and full-corpus training while using roughly 10--15\% of the full corpus. These exploratory findings illustrate a \emph{less is more} pattern and highlight how the value of code data in post training depends on which categories are combined for which model and task.

1 INTRODUCTION

Prior work examines code exposure and data properties, but does not specify which computational categories benefit a given model or how category benefits change under joint training. This study evaluates categories and compact mixtures, finding recurring QA gains alongside model- and task-dependent tradeoffs.

Motivation and approach

  • The study groups execution-verified code solutions by computational pattern to compare category-specific effects and test whether those benefits extend to joint training.Existing work on code exposure, representation, and complexity does not identify which computational categories benefit a given model or how their benefits change when combined.Source: §1 · pp. 1–2

RQ1

  • The balanced mixture improves average question-answering performance on all six backbones, while mathematics and code-generation effects vary across models.Positive task-family averages can coexist with declines on individual tasks.Source: §1 · p. 2

RQ2

  • The best-performing category varies with both the starting model and target task, and rankings can reverse when either changes.For every backbone, at least one single-category model has a higher evaluation-suite average than the example-matched balanced mixture.Source: §1 · p. 2

RQ3

  • Compact two- and three-category mixtures outperform both their best constituent and full-corpus training on selected model–task pairs while using roughly 10–15% of the full corpus.Each constituent also improves the target task when trained individually; the findings are exploratory and depend on the starting model and task.Source: §1 · p. 2

2 RELATED WORK

Related work studies code supervision and data selection from several angles, including exposure, data properties, example selection, and mixture optimization. This study focuses on computational patterns in reference solutions and tests category-specific transfer and within-code composition.

Code supervision

Data selection and composition

  • Data-selection work uses influence estimates, complexity, quality, diversity, semantic tags, or mixture models, while this study groups code by how reference solutions compute answers.The study compares category-specific transfer across models and tasks, then tests compact mixtures against individual constituents and full-corpus training.Source: §2 · p. 3

3 STUDY DESIGN: FROM CODE CATEGORIES TO COMPACT MIXTURES

The study decomposes code by the dominant computational pattern in each reference solution, measures transfer against the starting checkpoint, and recombines selected categories to examine joint-training outcomes.

3 STUDY DESIGN: FROM CODE CATEGORIES TO COMPACT MIXTURES

  • The design connects overall transfer, differences among categories, and the effects of training categories together.The ten-category taxonomy distinguishes algorithmic strategies and data-processing patterns, with the source-pool distribution summarized in Table 1.Source: §3 · p. 3, §3.1 · p. 3, Table 1 · p. 4

3.1 DECOMPOSING CODE BY COMPUTATIONAL PATTERN

  • Examples are assigned to ten non-overlapping primary categories based on the dominant computational pattern of the reference solution, rather than prompt subject matter.Auxiliary tags do not affect group assignment or training-subset selection, and overlapping patterns mean the groups do not isolate independent reasoning abilities.Source: §3.1 · p. 3

3.2 MEASURING CATEGORY-SPECIFIC TRANSFER

  • Transfer is the score change from the original checkpoint, measured in percentage points; positive values indicate improvement and negative values indicate decline.The study compares the ten-category balanced mixture with BASE for RQ1 and compares individual categories with one another and with that mixture for RQ2.Source: §3.2 · p. 4

3.3 RECOMBINING CATEGORIES INTO COMPACT MIXTURES

  • The study combines selected categories and compares joint training with individual constituents and full-corpus training to test whether individual-category benefits persist or extend.The screen requires every constituent to beat BASE and the joint model to beat both its best constituent and COMPLETE; it is exploratory because rankings and mixture evaluation use the same suite and cases are selected after observing results.Source: §3.3 · p. 4

4 EXPERIMENTAL SETUP

The experiments fine-tune six instruction-tuned backbones on single categories, balanced mixtures, and compact recipes, then evaluate eleven configurations spanning code generation, mathematics, and question answering.

Backbone models

  • Six instruction-tuned backbones are used, and every training run starts from its corresponding BASE checkpoint.The backbones are Qwen2.5-7B-Instruct, Qwen3-4B-Instruct-2507, Llama-3.1-8B-Instruct, Gemma-2-9B-Instruct, Llama-3.2-3B-Instruct, and Qwen1.5-7B-Chat.Source: §4 · p. 4

Training

  • For RQ1 and RQ2, each single-category configuration and the balanced ten-category reference use 1,706 examples and 107 optimization steps.Example counts and updates are matched across these configurations, but token counts are not; training uses LoRA supervised fine-tuning (Hu et al., 2022).Source: §4 · p. 5

Compact mixtures

  • Compact recipes use per-backbone Overall rankings for Personalized Top2 and Top3, while Shared Top3 uses the same three categories for all backbones.Recipes are fixed before mixture training; outcomes are then assessed task by task.Source: §4 · p. 5
  • Top3 recipes and their balanced and proportional all-category controls use 5,118 examples and 320 steps, while Top2 uses 3,412 examples and 214 steps.COMPLETE uses the full 35,974-example pool; differing example counts and token exposure mean these comparisons do not isolate gains caused by combining categories.Source: §4 · p. 5

Evaluation

  • Eleven configurations cover code generation, mathematics, and QA; evaluation reports three-seed mean scores and base-relative changes without significance tests.Code generation uses pass@1, other tasks use accuracy, and task-family averages weight configurations equally.Source: §4 · p. 5

5 RQ1: DOES CODE POST-TRAINING IMPROVE PERFORMANCE ACROSS TASKS AND MODELS?

Balanced-mixture fine-tuning raises average QA performance across all six backbones, but mathematics and code responses vary by model and individual tasks can decline. Single-category training also improves QA means broadly, while gains outside QA are less consistent.

Task-level variation

  • Positive task-family averages can conceal declines on individual tasks, so separate task results qualify aggregate gains.For example, the balanced mixture has four QA task declines despite improving the QA mean on every backbone.Source: §5 · p. 5

QA performance

  • The balanced mixture raises the five-task QA mean by +0.6 to +2.9 pp across all six backbones, improving 26 of 30 model–task scores.Four declines occur on Qwen3 FinQA, Llama-3.1 LegalBench, and Qwen1.5 LegalBench and MedCalc.Source: §5 · p. 5, Figure 2 · p. 6

Code and mathematics

  • The balanced mixture improves Code means on three backbones and Math means on four, but Qwen3 and Gemma-2 show opposing family-level changes.Qwen3 gains +1.4 pp in Math while losing 6.5 pp in Code; Gemma-2 gains +0.5 pp in Code while losing 3.0 pp in Math.Source: §5 · p. 5, Figure 3 · p. 6

Single-category transfer

  • All sixty single-category configurations improve the QA mean, while only 24 improve the combined code-and-mathematics mean.In a single-seed audit, all sixty retain positive five-task QA means, but only 45 remain positive after excluding ARC-Challenge and ScienceQA.Source: §5 · p. 6

6 RQ2: WHICH CODE CATEGORIES HELP WHICH TASKS ON WHICH MODELS?

Category-level responses depend on both the target task and starting model: rankings can reverse across tasks or backbones, and no category is universally preferred. Even recurring QA strengths coexist with task-specific declines.

Response map

  • Figure 4 maps 660 three-seed category–task changes from Base, with rows as categories and columns grouped by QA, math, and code.Rows within a task reveal category preferences; panel comparisons show their dependence on the backbone, while blue/red indicate gains/declines.Source: Figure 4 · p. 7

Task-dependent responses

  • A category’s effect can reverse across tasks: on Llama-3.1, implementation changes MATH-M by +7.6 pp while dynamic programming changes it by −2.9 pp, with the ordering reversed on MATH-H.On MATH-H, implementation changes the score by −1.1 pp and dynamic programming by +1.5 pp.Source: §6 · p. 7

Backbone-dependent responses

  • The same MATH-M comparison reverses across Llama backbones: implementation leads on Llama-3.1, while dynamic programming leads on Llama-3.2.On Llama-3.2, implementation lowers the score by 1.9 pp and dynamic programming improves it by +1.9 pp.Source: §6 · p. 7

Aggregate rankings

  • Aggregate category rankings vary by backbone, and each backbone has a single-category Overall score above its example-matched balanced mixture.The largest single-category margin ranges from +0.5 to +1.2 pp; examples of leading categories differ across Qwen2.5, Qwen3, and Llama-3.2.Source: §6 · pp. 7–8

Recurring strengths and limits

  • Implementation raises the QA-family mean on all six backbones, yet still lowers specific QA task scores on Llama-3.1 and Qwen1.5.It gives the largest single-category QA gain on Qwen2.5, Qwen3, and Llama-3.1, but declines on FinQA, LegalBench, or MedCalc in the stated cases.Source: §6 · p. 8

7 RQ3: CAN COMPACT MIXTURES OUTPERFORM SINGLE-CATEGORY AND FULL-CORPUS TRAINING?

Jointly trained compact mixtures outperform both their best constituent and full-corpus training on selected model–task pairs, using substantially fewer examples. These exploratory successes vary across models and tasks rather than establishing a general recipe.

Overall result

  • Compact mixtures outperform both their best constituent and full-corpus training on selected model–task pairs while using only 9.5–14.2% of the examples.The qualifying screen retains 9 of 66 two-category and 18 of 132 three-category comparisons, spanning five backbones and nine evaluation configurations.Source: §7 · p. 8

Representative results

  • On Qwen1.5 ARC-Challenge, the Math/Hash Top2 mixture reaches 57.76%, exceeding its best constituent by 0.60 pp and COMPLETE by 3.07 pp.It uses 3,412 examples, or 9.5% of the full corpus.Source: §7 · p. 8, Figure 5 · p. 8
  • On Qwen1.5 MBPP-Simple, the compact Imp+Hash+Sort mixture exceeds Complete, trained on all 35,974 examples, by 3.11 pp; it uses 5,118 examples, or 14.2% of Complete.Every constituent in both examples individually outperforms Base.Source: Figure 5 · p. 8, §7 · p. 9

Recipe variability

  • Successful recipes vary by model and task, and single-category rankings do not determine which jointly trained mixtures will perform best.The qualifying ARC-Challenge Top2 recipe combines Imp/Str on Qwen2.5 and Sort/Imp on Llama-3.1; matched-budget Top3 comparisons do not consistently beat balanced sampling on Overall.Source: §7 · p. 9

8 DISCUSSION

The discussion distinguishes category screening from mixture evaluation and cautions against treating aggregate gains or selected mixture wins as general rules. The evidence is bounded by the Python corpus and taxonomy, and the transfer mechanisms remain unresolved.

Implications for data selection

  • Category rankings can narrow candidate training data, but they do not determine joint-mixture performance or establish gains on a particular target task.Selection should consider both training-set composition and the distribution of gains across intended tasks.Source: §8 · p. 9

Scope of evidence

  • The observed responses are specific to execution-verified Python and the ten-category taxonomy, which groups solutions by dominant computational pattern.Broad groups such as implementation/utilities may contain finer distinctions; results are effects of defined subsets, not isolated reasoning abilities.Source: §8 · p. 9

Mixture-screen limitations

  • The mixture screen tests selected recipes, not a general selection rule, and does not establish optimal composition or causal synergy.Recipes were selected using the same evaluation suite, and constituent, mixture, and COMPLETE runs differ in training exposure; frozen recipes need independent matched-budget evaluation.Source: §8 · p. 9

Mechanisms

  • The causal mechanisms behind category-dependent transfer remain unestablished, and item-level audits do not quantify variability across training seeds.Audits identify candidate explanations, including possible changes in output behavior or evaluation interfaces.Source: §8 · p. 9

9 CONCLUSION

The conclusion presents category-level analysis and composition as a way to study code-data transfer and construct compact mixtures. The accompanying records document reported scores and the matched example and update counts used in training.

9 CONCLUSION

  • On selected model–task pairs, compact mixtures of individually beneficial categories outperform both their best constituent and full-corpus training while using roughly 10–15% of the full corpus.The authors describe these exploratory findings as a “less is more” pattern, with outcomes depending on the model and task.Source: §9 · p. 9

A SIX-MODEL RESULT RECORDS AND TRAINING PROTOCOL

  • The reported score records contain three-seed means, base-relative changes, and derived summaries, but do not recover hidden precision or per-seed variance.The records retain the precision exposed by the plots; independently printed changes agree with absolute-score differences within 0.1 pp.Source: A SIX-MODEL RESULT RECORDS AND… · p. 14

A.1 SAMPLING AND TRAINING CONFIGURATION

  • Each trained configuration uses 1,706 examples and 107 optimizer steps, while token totals are not matched across configurations.The balanced arm samples from the single-category subsets; Base receives no additional training.Source: §A.1 · p. 14, Table 2 · p. 14

A.2 EVALUATION CONFIGURATION AND RECORD COVERAGE

The section reports category-level and compact-mixture results across task families, backbones, and evaluation settings. It also documents evaluation and training choices, budget sensitivity, and limits on interpreting the reported means.

A.2 EVALUATION CONFIGURATION AND RECORD COVERAGE

  • Evaluation covers eleven code, mathematics, and QA configurations using task-specific scoring, while the math evaluator’s fallback matching retains parser compatibility limits.The additional audit record covers seed 20260730, but exact replication of the three-seed experiment still requires records for the other seeds and related training details.Source: §A.2 · p. 15

B.1 TASK-FAMILY SCORES FOR ALL TRAINING RECIPES

  • All 60 single-category configurations improve the QA mean, whereas only 24 improve the combined code-and-mathematics mean.Table 4 compares task-family means for Base, the balanced mixture, and all ten single-category recipes.Source: Table 4 · p. 16

B.2 AGGREGATE RESPONSES ACROSS BACKBONES

  • Single-category training improves Overall in all ten categories for Qwen2.5-7B but only four for Llama-3.1-8B; Qwen2.5 has the strongest aggregate gains across backbones.Direct implementation raises Qwen2.5 Overall from 55.8% to 58.7%; Llama-3.1 category responses range from −0.4 to +0.7 pp.Source: §B.2 · pp. 16–17

C COMPACT-MIXTURE EXPERIMENTS AND FURTHER VALIDATION

  • The shared three-category recipe combines implementation/utilities, hashing/sets, and sorting/ordered processing; personalized Top3 recipes use each backbone’s highest-ranked Overall categories.The shared recipe was selected heuristically, and the personalized recipes were frozen before mixture training but evaluated on the same configurations used for earlier selection.Source: C COMPACT-MIXTURE EXPERIMENTS AND… · p. 21, §C.1 · p. 21

C.2 OVERALL SCORES AND MATCHED-BUDGET COMPARISONS

  • 20 of 24 compact arms raise Overall and all 24 raise the five-task QA mean, but Qwen3 and Llama-3.1 favor Complete over their best compact recipe.The highest compact score is selected retrospectively from four candidates, not independently confirmed as an optimum.Source: §C.2 · p. 22, Table 6 · p. 22
  • Qwen2.5 Shared exceeds Complete by +11.429 pp on MATH Medium and +9.542 pp on MATH High, but trails Base by smaller margins and underperforms Complete on two code tasks.The complete task-score map displays the Qwen backbone comparisons; Overall is reported separately.Source: §C.2 · p. 22, Figure 13 · p. 24

C.3 GENERATION-BUDGET SENSITIVITY AND RESOURCE INTERPRETATION

  • At a 1,024-token GSM8K cap, 34 of 36 model–arm means increase, but positive trained-arm changes from same-budget Base fall from 16 of 30 to 6 of 30.For Qwen2.5 Shared, the advantage over Complete narrows from +2.426 pp at 256 tokens to about +0.15 pp at 1,024 tokens.Source: §C.3 · p. 23, Table 9 · p. 23

C.4 FURTHER CONFIRMATION

  • The complete task maps show six training arms across eleven configurations, with color indicating change from each backbone’s Base; independent validation is left for future work.The proposed validation would use separate selection data, held-out tasks or questions, paired seed records, and matched-budget controls.Source: §C.4 · p. 23, Figure 13 · p. 24, Figure 14 · p. 25

C.6 TASK-LEVEL SCREENING AGAINST CONSTITUENT CATEGORIES AND COMPLETE

The task-level screen finds compact mixtures that can exceed both individual constituents and full-corpus training, but the results are exploratory and do not establish category synergy. The accompanying audit finds consistent Base-relative gains for direct implementation on two QA benchmarks, while its proposed relational explanation is not supported by the sample.

C.6 TASK-LEVEL SCREENING AGAINST CONSTITUENT CATEGORIES AND COMPLETE

  • 27 stricter two- and three-category passes cover 16 distinct model–task pairs, five backbones, and nine evaluation configurations.The two-category screen retains 9/66 stricter passes; the three-category screen retains 18/132.Source: §C.6 · pp. 26–27
  • On ARC-C, Qwen1.5’s Top2 mixture reaches 57.76%, above both constituents and Complete (54.69%); on MBPP-Simple, Shared reaches 56.03%, versus Complete’s 52.92%.Both mixtures also exceed the larger Balanced and Proportional controls.Source: Table 10 · p. 26, §C.6 · p. 27, Table 11 · p. 27
  • The screen is exploratory: unequal token exposure means outperforming constituents does not isolate category synergy, and outperforming Complete does not establish optimality or statistical significance.Selections followed observation of results; independent recipe selection and matched-budget confirmation remain future work.Source: §C.6 · p. 27

D.1 AUDIT SCOPE AND COMPARATIVE GAINS

  • In the one-seed audit, direct implementation improves ARC-Challenge and ScienceQA over Base on every backbone, but its mean edge over the other nine categories has exceptions.The comparator is an arithmetic mean of independently trained arms, not an ensemble.Source: §D.1 · p. 28, Table 12 · p. 28

D.2 ILLUSTRATIVE CASES AND THE SCORING PROTOCOL

  • The audit’s selected answer changes are analyst interpretations, and the scoring protocol ranks complete candidate answers by mean token log probability without generating explanations.Thus, the observed gains cannot directly result from shorter generated explanations avoiding truncation or answer-extraction failures.Source: §D.2 · p. 28, Table 13 · p. 29

D.3 DO ADDITIONAL GAINS CONCENTRATE IN RELATIONAL QUESTIONS?

  • The sample does not support the prediction that direct implementation’s additional gains concentrate in relational questions: ARC-Challenge points the opposite way, and the ScienceQA interval spans zero.The comparison is exploratory and does not establish the absence of any rule-application benefit.Source: §D.3 · pp. 29–30, Table 14 · p. 29
  • On ScienceQA natural-science subsets, physical-relation and experimental-constraint questions have small relative advantages over other categories while declining versus Base.The total ScienceQA score also includes language- and social-science questions.Source: §D.3 · p. 30, Table 15 · p. 30

D.4 TRAINING CONTENT, RESPONSE STYLE, AND OPEN QUESTIONS

The direct-implementation audit reports task-dependent response changes and token exposure differences that limit interpretation. A separate QA audit finds positive single-category mean gains across the evaluated tasks, but those directions partly depend on benchmark composition.

D.4 TRAINING CONTENT, RESPONSE STYLE, AND OPEN QUESTIONS

  • Direct-implementation arms match 1,706 examples and 107 updates, but token exposure differs, leaving content, length, and supervision density coupled.For Qwen2.5, direct implementation uses 643,839 input and 208,684 supervised tokens, compared with 772,561/284,522 for dynamic programming and 980,483/326,472 for graph training.Source: §D.4 · p. 30
  • For Qwen3, mean visible GSM8K length falls from 181.5 to 163.7 tokens, while FinQA rises from 40.2 to 43.8 and MedCalc from 39.4 to 51.7.Selected GSM8K cases show changed solution steps, but compare only Base with direct implementation and cannot establish an advantage over other categories.Source: §D.4 · p. 30

E AUDIT OF QA COVERAGE AND BACKBONE-DEPENDENT RESPONSES

  • The appendix’s single-seed audit covers saved outputs and protocol records but does not estimate training variability or isolate training-content effects from all other variation.The main reported results retain three-seed aggregates.Source: E AUDIT OF QA COVERAGE AND… · p. 31, §E.1 · p. 31

E.2 HOW BROADLY DO THE QA GAINS EXTEND?

  • All sixty single-category configurations improve the five-task QA mean in the audited seed, but only 35 improve every constituent task individually.Removing ARC-Challenge and ScienceQA leaves 45 positive comparisons.Source: §E.2 · p. 31, Table 16 · p. 31
  • Removing ARC-Challenge alone leaves 57/60 positive comparisons, and removing ScienceQA alone leaves 56/60, so the universal five-task direction partly depends on task composition.The remaining positive transfer does not establish uniform gains across QA domains or a particular reasoning mechanism.Source: §E.2 · p. 31

E.3 GSM8K: SHORTER RESPONSES ACCOMPANY OPPOSITE SCORE CHANGES

The audit finds that all categories shorten GSM8K responses even as score directions split by backbone, and that code-generation failures can reflect interface changes as well as program behavior. Its selected cases and records do not establish population-level causes.

E.3 GSM8K: SHORTER RESPONSES ACCOMPANY OPPOSITE SCORE CHANGES

  • Every category improves GSM8K for Qwen2.5, Qwen3, and Llama-3.1, but lowers it for Gemma-2, Llama-3.2, and Qwen1.5; all sixty arms also have shorter visible responses than Base.Output shortening alone therefore does not distinguish positive from negative score changes.Source: §E.3 · p. 32, Table 17 · p. 32
  • Replaying the archived GSM8K scoring rule reproduces all 94,968 saved correctness flags, while an alternative answer-marker sensitivity check changes 440 originally correct records to incorrect and restores none.The alternative is not a validated correction or evidence that extraction errors are absent.Source: §E.3 · p. 32

E.4 CODE FAILURES INCLUDE SEMANTIC AND INTERFACE CHANGES

  • On HumanEval, prompt-helper omissions and wrapper indentation can contribute to failed tests, but static syntax repair was not executed and cannot count as a recovered correct answer.All original test outcomes remain unchanged.Source: §E.4 · p. 32, Table 18 · p. 32
  • Shorter HumanEval outputs accompany both 150/154 Qwen3 regressions and 33/36 improvements, so brevity alone cannot explain correctness changes.Gemma-2’s mean visible length instead rises from 67.5 to 87.1 tokens while extracted-code character count falls.Source: §E.4 · p. 33

E.5 ILLUSTRATIVE CASES AND INTERPRETATION LIMITS

  • The case audit uses fixed samples with one AI reviewer and does not estimate error-type prevalence; its observations do not identify population-level causal contributions.Original generation token IDs and stopping reasons are unavailable, and the selected short-output subsets cannot establish an effect with truncation held constant.Source: §E.5 · p. 33, Table 19 · p. 33
Loading 2609.33845v1…