arxiv: 2604.13413 · v1 · submitted 2026-04-15 · 💻 cs.LG

Recognition: unknown

Dataset-Level Metrics Attenuate Non-Determinism: A Fine-Grained Non-Determinism Evaluation in Diffusion Language Models

Zhengyu Fang , Zhimeng Jiang , Huiyuan Chen , Xiaoge Zhang , Tianyi Li , Kaiyu Tang , Xiao Li , Jing Li

Authors on Pith no claims yet

Pith reviewed 2026-05-10 13:28 UTC · model grok-4.3

classification 💻 cs.LG

keywords diffusion language modelsnon-determinism evaluationdataset-level metricssample-level analysisFactor Variance Attributioncode generationquestion answering

0 comments

The pith

Dataset-level metrics systematically mask non-determinism in diffusion language models by averaging across samples and runs.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that standard dataset-level metrics for diffusion language models hide substantial variability because they combine prediction quality from many individual samples and separate runs into single aggregate scores. This averaging produces the appearance of stability even when the same model and input produce different outputs under small changes to guidance scale, diffusion steps, batch size, hardware, or numerical precision. Shifting attention to direct sample-level prediction differences uncovers that non-determinism is both widespread and patterned, appearing far more strongly in code-generation tasks than in question answering. The authors introduce Factor Variance Attribution to measure how much of the observed variation can be traced to each factor setting.

Core claim

Dataset-level metrics attenuate non-determinism in diffusion language models by aggregating sample-level prediction quality across different runs. As a result, configurations with similar aggregate performance can exhibit substantially different behaviors on individual inputs. Fine-grained analysis across model and system factors reveals that non-determinism is pervasive and structured, with code generation showing markedly higher sensitivity than question answering.

What carries the argument

Factor Variance Attribution (FVA), a cross-factor metric that decomposes observed sample-level non-determinism into portions attributable to each evaluation factor such as guidance scale or hardware precision.

If this is right

Configurations that look equivalent under dataset averages can differ substantially on specific inputs and error types.
Non-determinism strength depends on task type, reaching higher levels in code generation than in question answering.
Both model factors such as diffusion steps and system factors such as numerical precision contribute measurable variance.
Reliable assessment of diffusion language models requires factor-aware, sample-level metrics in addition to aggregates.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Reporting only averages may lead practitioners to underestimate deployment risk when outputs must remain consistent on single queries.
The same sample-level decomposition could be applied to other generative architectures to locate common sources of run-to-run drift.
Developers could add variance-aware benchmarks that require models to keep individual-sample disagreement below a stated threshold.

Load-bearing premise

Observed differences in sample-level predictions across runs and factor settings reflect the models' inherent non-determinism rather than unaccounted implementation details or limited Monte Carlo sampling.

What would settle it

Reproduce the sample-level comparisons while holding every implementation detail fixed, using identical hardware, and drawing thousands of Monte Carlo samples per input; if prediction differences across runs then disappear or fall within expected statistical noise, the claim of pervasive structured non-determinism would be falsified.

Figures

Figures reproduced from arXiv: 2604.13413 by Huiyuan Chen, Jing Li, Kaiyu Tang, Tianyi Li, Xiaoge Zhang, Xiao Li, Zhengyu Fang, Zhimeng Jiang.

**Figure 1.** Figure 1: Sample-level prediction flip rates across inference-time configurations. Each panel varies one inference factor while holding all other factors fixed: (a) batch size, (b) classifier-free guidance (CFG) scale, and (c) number of Monte Carlo (MC) samples. For a given sample s, the prediction flip rate is defined as 1 − maxa|{c ∈ C : ˆys,c = a}|/|C|, where yˆs,c denotes the predicted label for sample s under c… view at source ↗

**Figure 2.** Figure 2: This hierarchy underlies our sample-aware and factor-aware evaluation paradigm and is formally instantiated by the definitions in Section 3.2. differ (Ribeiro et al., 2020). The consistency of this phenomenon across factors suggests that such non-determinism is not an isolated artifact, but a recurring property of diffusion language model evaluation. Towards Factor-Aware and Sample-Aware Evaluation. The … view at source ↗

**Figure 3.** Figure 3: Within-factor variability across datasets and backbones. The figure reports the standard deviation (Std) of evaluation scores across different settings within each factor, aggregated by dataset. Larger Std indicates stronger sensitivity to specific settings of the same factor, reflecting pronounced setting-level non-determinism. Results are shown for both LLaDA and LLaDA-1.5 on question answering and code … view at source ↗

**Figure 4.** Figure 4: Uncertainty of factor-level effects across datasets and backbones. The figure reports the standard error (SE) of the mean evaluation score for each factor, aggregated across its settings. SE reflects the reliability of the estimated factor-level effect and complements within-factor variability by characterizing uncertainty at the factor level. Together with FVA, this analysis clarifies how evaluation non-d… view at source ↗

read the original abstract

Diffusion language models (DLMs) have emerged as a promising paradigm for large language models (LLMs), yet the non-deterministic behavior of DLMs remains poorly understood. The existing non-determinism evaluations for LLMs predominantly rely on dataset-level metrics under fixed inference configurations, providing limited insight into how model behavior varies across runs and evaluation conditions. In this work, we show that dataset-level metrics systematically attenuate non-determinism in diffusion language models by aggregating sample-level prediction quality across different runs. As a result, configurations with similar aggregate performance can exhibit substantially different behaviors on individual inputs, leaving fine-grained instability and distinct error patterns uncharacterized. To address this limitation, we conduct a fine-grained evaluation of non-determinism based on sample-level prediction differences across a range of model-related factors-including guidance scale, diffusion steps, and Monte Carlo sampling-as well as system-related factors such as batch size, hardware, and numerical precision. Our analysis reveals that non-determinism in DLMs is pervasive and structured, with code generation exhibiting markedly higher sensitivity to factor-level choices than question answering. To attribute sources of non-determinism evaluation, we introduce Factor Variance Attribution (FVA), a cross-factor analysis metric that decomposes observed non-determinism into variance attributable to different evaluation factor settings. Our findings highlight the need for fine-grained, factor-aware evaluation to enable reliable non-determinism assessment of diffusion language models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Dataset-level metrics hide per-sample instability in diffusion LMs, and FVA decomposes the variance by factor, but the measurements need tighter controls on implementation noise.

read the letter

The main thing to know is that dataset-level scores for diffusion language models average away real differences in how the model behaves on individual examples across runs. Two configurations can look similar in aggregate but produce distinct error patterns on the same inputs, and the paper shows this happens systematically when you switch guidance scale, diffusion steps, or Monte Carlo samples. They also vary system factors like batch size, hardware, and precision, which is a useful addition. Code generation turns out more sensitive than question answering, which tracks with the structured nature of the output. Factor Variance Attribution then splits the observed variation across those factors, giving a concrete way to see which settings drive the instability. This is new relative to most prior non-determinism work on LLMs, which sticks to fixed configs and overall metrics. The empirical sweep is straightforward and the decomposition metric is easy to apply, so the core observation stands on its own. The soft spot is whether the sample-level differences are purely model stochasticity or partly experimental artifacts. The stress-test note is right to flag things like per-step RNG state, floating-point reduction order, or library nondeterminism flags. If those were not locked down tightly, some of the attributed variance could be noise rather than inherent behavior. The paper does vary system factors on purpose, but without explicit controls or deterministic-mode checks the numbers may overestimate the model-driven part. This is worth a serious referee for groups working on diffusion generation or reliability testing. The claim is clear, the metric is reproducible, and the gap it identifies is real even if the exact percentages need cleaner isolation. I would send it for review with a request to document the random-state handling in more detail.

Referee Report

1 major / 1 minor

Summary. The paper claims that dataset-level metrics attenuate non-determinism in diffusion language models (DLMs) by averaging sample-level prediction quality across runs, such that configurations with similar aggregates can hide substantially different per-input behaviors. It conducts a fine-grained analysis of non-determinism across model factors (guidance scale, diffusion steps, Monte Carlo sampling) and system factors (batch size, hardware, numerical precision), reports that non-determinism is pervasive and task-dependent (higher in code generation than QA), and introduces Factor Variance Attribution (FVA) as a cross-factor decomposition metric to attribute observed variance to specific evaluation settings.

Significance. If the empirical findings hold, the work identifies a systematic limitation in how non-determinism is currently assessed for an emerging LLM paradigm, motivating factor-aware and sample-level evaluation protocols. The FVA metric offers a concrete tool for variance decomposition that could be adopted more broadly, and the task-specific sensitivity results provide actionable guidance for inference configuration choices in DLMs.

major comments (1)

[Abstract and §3] Abstract and §3 (Evaluation Methodology): The central attenuation claim requires that observed sample-level prediction differences across runs and factor settings primarily reflect inherent model non-determinism rather than unaccounted implementation artifacts. The manuscript states that system factors are evaluated, yet provides no explicit controls or reporting for diffusion-specific sources of nondeterminism such as per-denoising-step RNG state, floating-point reduction order in attention, or library nondeterminism flags; without these, the base measurements feeding FVA may be contaminated, undermining the attribution and the claim that dataset-level metrics systematically attenuate true non-determinism.

minor comments (1)

[Abstract] The abstract refers to 'Monte Carlo sampling' as a factor but does not specify the number of samples drawn per configuration or the exact variance estimator used; this detail should be added in the methods for reproducibility.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and for identifying a key methodological gap. We agree that stronger documentation of controls is needed to support the central claims. We address the major comment below and will revise the manuscript accordingly.

read point-by-point responses

Referee: [Abstract and §3] Abstract and §3 (Evaluation Methodology): The central attenuation claim requires that observed sample-level prediction differences across runs and factor settings primarily reflect inherent model non-determinism rather than unaccounted implementation artifacts. The manuscript states that system factors are evaluated, yet provides no explicit controls or reporting for diffusion-specific sources of nondeterminism such as per-denoising-step RNG state, floating-point reduction order in attention, or library nondeterminism flags; without these, the base measurements feeding FVA may be contaminated, undermining the attribution and the claim that dataset-level metrics systematically attenuate true non-determinism.

Authors: We acknowledge that the manuscript provides insufficient explicit reporting on diffusion-specific nondeterminism sources. Although system factors (batch size, hardware, numerical precision) were varied and reported at a high level, we did not document per-denoising-step RNG state handling, attention reduction orders, or library flags such as torch.use_deterministic_algorithms. In the revision we will expand §3 (Evaluation Methodology) with: (i) the precise RNG seeding protocol applied at each diffusion step, (ii) the nondeterminism flags and environment settings used, and (iii) any steps taken to stabilize floating-point reduction order. These additions will allow readers to assess whether the sample-level differences primarily reflect the intended model and system factors. We maintain that the observed attenuation of non-determinism by dataset-level metrics remains valid across the tested configurations, but we agree that the requested documentation is necessary to fully substantiate the FVA attribution. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical definitions and variance decomposition remain independent of inputs

full rationale

The paper is an empirical evaluation that defines non-determinism directly from observed sample-level prediction differences across runs and factor settings, then reports the attenuation effect of dataset-level aggregation as a measured outcome rather than a derived equation. The introduced FVA metric decomposes observed variance into factor contributions via standard cross-factor analysis without reducing to fitted parameters or prior self-citations. No load-bearing steps invoke self-definitional loops, uniqueness theorems from the same authors, or ansatzes smuggled via citation; the central claims rest on experimental measurements that can be independently replicated or falsified.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that sample-level prediction differences across runs constitute measurable non-determinism and that the chosen factors are the relevant sources of variance.

axioms (1)

domain assumption Non-determinism in DLMs can be quantified by differences in per-sample predictions across repeated runs under varied factor settings
Invoked when the paper defines fine-grained evaluation and attributes variance to factors.

invented entities (1)

Factor Variance Attribution (FVA) no independent evidence
purpose: Decompose observed non-determinism into contributions from individual evaluation factors
New metric introduced to attribute sources of instability

pith-pipeline@v0.9.0 · 5585 in / 1260 out tokens · 50463 ms · 2026-05-10T13:28:36.942538+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

31 extracted references · 28 canonical work pages · 9 internal anchors

[1]

I., Licorish, S., Wang, F., and Treude, C

Arora, C., Sayeed, A. I., Licorish, S., Wang, F., and Treude, C. Optimizing large language model hyperparameters for code generation.arXiv preprint arXiv:2408.10577,

work page arXiv
[2]

Program Synthesis with Large Language Models

Austin, J., Johnson, D. D., Ho, J., Tarlow, D., and Van Den Berg, R. Structured denoising diffusion models in discrete state-spaces.Advances in neural information processing systems, 34:17981–17993, 2021a. Austin, J., Odena, A., Nye, M., Bosma, M., Michalewski, H., Dohan, D., Jiang, E., Cai, C., Terry, M., Le, Q., et al. Program synthesis with large langu...

work page internal anchor Pith review Pith/arXiv arXiv
[3]

DeepSeek LLM: Scaling Open-Source Language Models with Longtermism

Bi, X., Chen, D., Chen, G., Chen, S., Dai, D., Deng, C., Ding, H., Dong, K., Du, Q., Fu, Z., et al. Deepseek llm: Scaling open-source language models with longtermism. arXiv preprint arXiv:2401.02954,

work page internal anchor Pith review arXiv
[4]

D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. Language models are few-shot learners. Advances in neural information processing systems, 33: 1877–1901,

1901
[5]

Evaluating Large Language Models Trained on Code

Chen, M. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374,

work page internal anchor Pith review Pith/arXiv arXiv
[6]

Sdar: A syn- ergistic diffusion-autoregression paradigm for scalable sequence generation.arXiv preprint arXiv:2510.06303,

Cheng, S., Bian, Y ., Liu, D., Zhang, L., Yao, Q., Tian, Z., Wang, W., Guo, Q., Chen, K., Qi, B., et al. Sdar: A syn- ergistic diffusion-autoregression paradigm for scalable sequence generation.arXiv preprint arXiv:2510.06303,

work page arXiv
[7]

Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge

Clark, P., Cowhey, I., Etzioni, O., Khot, T., Sabharwal, A., Schoenick, C., and Tafjord, O. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457,

work page internal anchor Pith review Pith/arXiv arXiv
[8]

H., Doucet, A., Strudel, R., Dyer, C., Durkan, C., et al

Dieleman, S., Sartran, L., Roshannai, A., Savinov, N., Ganin, Y ., Richemond, P. H., Doucet, A., Strudel, R., Dyer, C., Durkan, C., et al. Continuous diffusion for categorical data.arXiv preprint arXiv:2211.15089,

work page arXiv
[9]

Studying how configurations impact code gener- ation in llms: The case of chatgpt.arXiv preprint arXiv:2502.17450,

Donato, B., Mariani, L., Micucci, D., and Riganelli, O. Studying how configurations impact code gener- ation in llms: The case of chatgpt.arXiv preprint arXiv:2502.17450,

work page arXiv
[10]

Diffuseq: Se- quence to sequence text generation with diffusion models.arXiv preprint arXiv:2210.08933,

Gong, S., Li, M., Feng, J., Wu, Z., and Kong, L. Diffuseq: Sequence to sequence text generation with diffusion mod- els.arXiv preprint arXiv:2210.08933,

work page arXiv
[11]

Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt

Gong, S., Zhang, R., Zheng, H., Gu, J., Jaitly, N., Kong, L., and Zhang, Y . Diffucoder: Understanding and improving masked diffusion models for code generation.arXiv preprint arXiv:2506.20639,

work page arXiv
[12]

20251026

doi: 10.64434/tml. 20250910. https://thinkingmachines.ai/blog/defeating- nondeterminism-in-llm-inference/. He, Z., Sun, T., Tang, Q., Wang, K., Huang, X.-J., and Qiu, X. Diffusionbert: Improving generative masked language models with diffusion models. InProceedings of the 61st annual meeting of the association for computational lin- guistics (volume 1: Lo...

work page doi:10.64434/tml
[13]

Measuring determinism in large language models for software code review.arXiv preprint arXiv:2502.20747,

Klishevich, E., Denisov-Blanch, Y ., Obstbaum, S., Ciobanu, I., and Kosinski, M. Measuring determinism in large language models for software code review.arXiv preprint arXiv:2502.20747,

work page arXiv
[14]

Mercury: Ultra- fast language models based on diffusion.arXiv preprint arXiv:2506.17298,

Labs, I., Khanna, S., Kharbanda, S., Li, S., Varma, H., Wang, E., Birnbaum, S., Luo, Z., Miraoui, Y ., Palrecha, A., et al. Mercury: Ultra-fast language models based on diffusion. arXiv preprint arXiv:2506.17298,

work page arXiv
[15]

Lavida-o: Elastic large masked diffusion models for unified multimodal understanding and generation

Li, S., Gu, J., Liu, K., Lin, Z., Wei, Z., Grover, A., and Kuen, J. Lavida-o: Elastic large masked diffusion models for unified multimodal understanding and generation.arXiv preprint arXiv:2509.19244, 2025a. Li, T., Chen, M., Guo, B., and Shen, Z. A survey on diffu- sion language models.arXiv preprint arXiv:2508.10875, 2025b. Li, X., Thickstun, J., Gulraj...

work page arXiv
[16]

Holistic Evaluation of Language Models

Liang, P., Bommasani, R., Lee, T., Tsipras, D., Soylu, D., Yasunaga, M., Zhang, Y ., Narayanan, D., Wu, Y ., Kumar, A., et al. Holistic evaluation of language models.arXiv preprint arXiv:2211.09110,

work page internal anchor Pith review Pith/arXiv arXiv
[17]

Kaijing Ma, Xinrun Du, Yunran Wang, Haoran Zhang, Zhoufutu Wen, Xingwei Qu, Jian Yang, Jiaheng Liu, Minghao Liu, Xiang Yue, et al

Liu, J., Dong, X., Ye, Z., Mehta, R., Fu, Y ., Singh, V ., Kautz, J., Zhang, C., and Molchanov, P. Tidar: Think in diffusion, talk in autoregression.arXiv preprint arXiv:2511.08923, 2025a. Liu, X., Song, Y ., Liu, Z., Huang, Z., Guo, Q., He, Z., and Qiu, X. Longllada: Unlocking long context capabilities in diffusion llms.arXiv preprint arXiv:2506.14429, 2...

work page arXiv
[18]

Large Language Diffusion Models

Nie, S., Zhu, F., You, Z., Zhang, X., Ou, J., Hu, J., Zhou, J., Lin, Y ., Wen, J.-R., and Li, C. Large language diffusion models.arXiv preprint arXiv:2502.09992,

work page internal anchor Pith review arXiv
[19]

Tess 2: A large-scale generalist diffusion language model.arXiv preprint arXiv:2502.13917,

Tae, J., Ivison, H., Kumar, S., and Cohan, A. Tess 2: A large-scale generalist diffusion language model.arXiv preprint arXiv:2502.13917,

work page arXiv
[20]

Llama 2: Open Foundation and Fine-Tuned Chat Models

Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y ., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al. Llama 2: Open foundation and fine- tuned chat models.arXiv preprint arXiv:2307.09288,

work page internal anchor Pith review Pith/arXiv arXiv
[21]

Wang, J. J. and Wang, V . X. Assessing consistency and reproducibility in the outputs of large language models: Evidence across diverse finance and accounting tasks. arXiv preprint arXiv:2503.16974,

work page arXiv
[22]

Lumina- dimoo: An omni diffusion large language model for multi- modal generation and understanding.arXiv preprint arXiv:2510.06308,

Xin, Y ., Qin, Q., Luo, S., Zhu, K., Yan, J., Tai, Y ., Lei, J., Cao, Y ., Wang, K., Wang, Y ., et al. Lumina- dimoo: An omni diffusion large language model for multi- modal generation and understanding.arXiv preprint arXiv:2510.06308,

work page arXiv
[23]

Qwen3 Technical Report

Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025a. Yang, L., Tian, Y ., Li, B., Zhang, X., Shen, K., Tong, Y ., and Wang, M. Mmada: Multimodal large diffusion language models.arXiv preprint arXiv:2505.15809, 2025b. Ye, J., Xie, Z., Zheng, L., ...

work page internal anchor Pith review Pith/arXiv arXiv
[24]

Llada-v: Large language diffusion models with visual instruction tuning.arXiv preprint arXiv:2505.16933, 2025

You, Z., Nie, S., Zhang, X., Hu, J., Zhou, J., Lu, Z., Wen, J.-R., and Li, C. Llada-v: Large language diffusion models with visual instruction tuning.arXiv preprint arXiv:2505.16933,

work page arXiv
[25]

Benchmarking large language model volatility.arXiv preprint arXiv:2311.15180,

Yu, B. Benchmarking large language model volatility.arXiv preprint arXiv:2311.15180,

work page arXiv
[26]

Dimple: Discrete diffusion multimodal large language model with parallel decoding

Yu, R., Ma, X., and Wang, X. Dimple: Discrete diffusion multimodal large language model with parallel decoding. arXiv preprint arXiv:2505.16990,

work page arXiv
[27]

Seqdif- fuseq: Text diffusion with encoder-decoder transformers

Yuan, H., Yuan, Z., Tan, C., Huang, F., and Huang, S. Seqdif- fuseq: Text diffusion with encoder-decoder transformers. arXiv preprint arXiv:2212.10325,

work page arXiv
[28]

Understanding and mitigating numerical sources of nondeterminism in llm inference

Yuan, J., Li, H., Ding, X., Xie, W., Li, Y .-J., Zhao, W., Wan, K., Shi, J., Hu, X., and Liu, Z. Understanding and mitigating numerical sources of nondeterminism in llm inference. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025a. Yuan, J., Zhang, J., Wen, A., and Hu, X. The sci- ence of evaluating foundation models.arXi...

work page arXiv
[29]

LLaDA 1.5: Variance-Reduced Preference Optimization for Large Language Diffusion Models

Zhu, F., Wang, R., Nie, S., Zhang, X., Wu, C., Hu, J., Zhou, J., Chen, J., Lin, Y ., Wen, J.-R., et al. Llada 1.5: Variance- reduced preference optimization for large language diffu- sion models.arXiv preprint arXiv:2505.19223,

work page internal anchor Pith review arXiv
[30]

Model outputs are evaluated based on functional correctness using hidden unit tests

5: A code generation benchmark composed of Python programming problems. Model outputs are evaluated based on functional correctness using hidden unit tests. • MBPP(Austin et al., 2021b) 6: The Mostly Basic Programming Problems dataset, which includes a larger collection of programming tasks with varying difficulty levels. Similar to HumanEval, evaluation ...

2015
[31]

Base” denotes the original configuration, while “Alt

Backbone Factor HumanEval MBPP Pass@1 Dataset-level Sample-level Pass@1 Dataset-level Sample-level LLaMA-2-7B Precision (Std)0.0650 0.0254 0.1517 0.1940 0.0140 0.3227 Precision (SE)0.0650 0.0147 0.0118 0.1940 0.0081 0.0144 Batch size (Std)0.0549 0.0122 0.1358 0.1935 0.0090 0.3172 Batch size (SE)0.0549 0.0061 0.0106 0.1935 0.0045 0.0142 Temperature (Std)0....

1940