arxiv: 2605.09060 · v1 · submitted 2026-05-09 · 💻 cs.CL

Recognition: 1 theorem link

· Lean Theorem

Language-Conditioned Visual Grounding with CLIP Multilingual

J. de Curt\`o , Mauro Liz , I. de Zarz\`a

Authors on Pith no claims yet

Pith reviewed 2026-05-12 02:12 UTC · model grok-4.3

classification 💻 cs.CL

keywords multilingual CLIPvisual groundinglow-resource languagestext encoderspatial misalignmentcross-language performanceXLM-RoBERTaCLIP probe

0 comments

The pith

Low-resource languages incur a text-branch penalty in multilingual CLIP visual grounding that persists when the visual encoder is held fixed.

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

The paper tests whether multilingual performance gaps in vision-language models originate in the visual encoder, the text branch, or their interaction. By fixing the visual encoder across thirteen languages and varying only the XLM-RoBERTa text branch, it isolates a structural deficit for low-resource languages such as Arabic, Basque, and Luxembourgish. This deficit appears at both base and large visual scales, with scaling widening gaps for some languages while narrowing them for others. The dominant failure mode is spatial misalignment rather than collapse of overall similarity signals. The findings indicate that equitable multilingual grounding requires targeted fixes to text processing rather than uniform visual scaling.

Core claim

Holding the visual encoder identical across languages while varying only the text branch reveals that low-resource languages incur a structural penalty in language-conditioned visual grounding at both backbone scales, with cluster-mask IoU gaps of 0.114 at base and 0.143 at large. Scaling the visual encoder 7x separates corpus-coverage failures from tokeniser-fertility failures and preserves peak similarity (mean ratio 0.94) while cluster-mask IoU drops, identifying spatial misalignment as the main issue rather than signal loss.

What carries the argument

Dense multilingual CLIP probe that keeps the visual encoder (ViT-B/32 or ViT-H/14) fixed and varies only the XLM-RoBERTa text branch, quantified by cluster-mask IoU, top-percentile IoU, and Spearman correlation on 11 concepts and 210 images.

If this is right

Targeted improvements to the text encoder can reduce grounding gaps for low-resource languages without retraining the visual backbone.
Scaling the visual encoder alone does not close all multilingual disparities and can increase some gaps.
Spatial misalignment dominates over signal collapse, so localization accuracy should be prioritized in multilingual training.
Energy costs of 3.4-3.9 Wh per 1,000 queries make repeated dense probing practical for comparing language performance.

Where Pith is reading between the lines

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

The same fixed-visual probe could be applied to other vision-language models to test whether text-branch isolation is a general pattern.
Language-specific tokenization or adapter layers on the text encoder might mitigate the observed penalties.
The results suggest that balanced coverage of low-resource languages in pretraining data is needed to prevent structural alignment disadvantages.

Load-bearing premise

That fixing the visual encoder and applying the chosen IoU and correlation metrics cleanly isolates text-branch effects without confounding from tokenization choices, training data overlap, or metric sensitivity.

What would settle it

If the same performance gaps appear when the text branch is fixed and only the visual encoder is varied across languages, that would show the deficits are not isolated to the text branch.

Figures

Figures reproduced from arXiv: 2605.09060 by I. de Zarz\`a, J. de Curt\`o, Mauro Liz.

**Figure 2.** Figure 2: Per-language IoU shift under scaling, ∆ (ℓ) IoU = IoU(ℓ) large − IoU(ℓ) base. Hatched bars mark low-resource languages. Basque (∆ = −0.056) and Luxembourgish (∆ = −0.076) lose cross-language agreement under scaling; Arabic (∆ = +0.033) and Chinese (Simp., ∆ = +0.039) recover. generalise better across distributional shifts in the input. The data falsify this prediction in the specific direction that matters… view at source ↗

**Figure 4.** Figure 4: Per-concept cluster-mask IoU for the three low-resource languages [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

**Figure 5.** Figure 5: places the probe on a uniform energy scale alongside autoregressive multilingual VLM budgets reported under the same NVML protocol [16], [17]. At 3.4–3.9 Wh per 1,000 queries, dense-CLIP grounding sits roughly 20–50× below typical generative VLM inference, with margin sufficient to absorb the cross-language consistency check proposed above. The substrate is not a substitute for generative VLMs in tasks th… view at source ↗

read the original abstract

Multilingual vision-language models exhibit systematic performance gaps across languages, but the mechanism remains ambiguous: cross-language divergence could arise from the visual encoder, the text branch, or their interaction. We resolve this ambiguity through a dense multilingual CLIP probe in which the visual encoder is held identical across thirteen typologically diverse languages and only the XLM-RoBERTa text branch varies. We evaluate two CLIP architectures spanning a 7x visual-encoder scale gap (XLM-R base + ViT-B/32, ~87M visual parameters; XLM-R large + ViT-H/14, ~632M) on 11 concepts and 210 images, and quantify cross-language agreement via cluster-mask IoU, top-percentile IoU, and Spearman rank correlation against an English reference (n=2,310 paired observations per language). Three findings emerge. First, low-resource languages (Arabic, Basque, Luxembourgish) incur a structural penalty at both backbone scales (Wilcoxon HR>LR p<10^-300; cluster-mask IoU gap +0.114 at base, +0.143 at large), isolating the deficit to the text branch. Second, scaling the encoder 7x widens the gap for structural failure cases (Basque {\Delta}=-0.056, Luxembourgish {\Delta}=-0.076) while improving Arabic ({\Delta}=+0.033), separating corpus-coverage from tokeniser-fertility failures. Third, peak similarity is preserved across languages (mean ratio 0.94 at large scale) while cluster-mask IoU drops sharply, identifying spatial misalignment, not signal collapse, as the dominant failure mode. At 3.4-3.9 Wh per 1,000 queries, dense-CLIP grounding is competitive with high-throughput inference budgets, positioning it as a practical substrate for energy-aware multilingual deployment.

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

The paper shows multilingual CLIP grounding gaps come from the text branch, with low-resource languages suffering spatial misalignment even after scaling.

read the letter

The main thing to know is that this work isolates the text encoder as the source of poor performance on low-resource languages in vision-language models. By freezing the visual encoder and feeding the same images to different language inputs through XLM-RoBERTa, they measure grounding quality on 11 concepts across 13 languages and find consistent drops for Arabic, Basque, and Luxembourgish at both model scales. The gaps show up in cluster-mask IoU and hold across Spearman correlation and top-percentile measures, with the failure mode being misalignment of attention rather than overall weak similarity scores. Scaling the visual encoder 7x helps Arabic but hurts Basque and Luxembourgish further, which separates tokenization issues from data coverage problems. The energy cost per query is reported as a practical aside but is not central to the claims. This controlled probe with fixed vision and quantitative region overlap metrics is the clearest new element, and the results line up across two architectures and thousands of paired observations. The design avoids circularity by benchmarking everything against English references computed the same way. The main limitation is the narrow scope: only 210 images and 11 concepts, which leaves open whether the patterns would hold for more varied scenes or additional languages. The abstract also skips details on how images were chosen or filtered, so it is hard to rule out selection effects without the full methods. Tokenization fertility is treated as intrinsic to the text branch, which is reasonable, but interactions with pretraining data overlap could still play a role not fully separated here. Readers working on multilingual model diagnostics or fairness checks will get the most out of it. The isolation approach and scale comparisons are solid enough to merit peer review, though referees will likely ask for more concepts, explicit controls for tokenization, and raw data release to verify the Wilcoxon results.

Referee Report

2 major / 2 minor

Summary. The paper claims that multilingual CLIP models exhibit performance gaps for low-resource languages that can be isolated to the text branch by holding the visual encoder (ViT-B/32 or ViT-H/14) fixed while varying only the XLM-RoBERTa text encoder across 13 languages. Using a probe on 11 concepts and 210 images with metrics including cluster-mask IoU, top-percentile IoU, and Spearman rank correlation against an English reference (2310 paired observations per language), the authors report structural penalties for Arabic, Basque, and Luxembourgish at both scales, differential scaling effects, and spatial misalignment (rather than signal collapse) as the dominant failure mode, while noting competitive energy use of 3.4-3.9 Wh per 1000 queries.

Significance. If the results hold, the work is significant for providing a controlled empirical isolation of text-branch deficits in multilingual vision-language models, with clear implications for tokenizer design and pretraining data balancing in low-resource settings. The design of fixing the visual encoder, using multiple complementary metrics, two backbone scales, and consistent statistical findings (e.g., Wilcoxon p<10^-300) adds robustness; the energy analysis further supports practical utility. This strengthens the case for targeted improvements in multilingual VLMs without requiring full model retraining.

major comments (2)

[Evaluation] Evaluation section: the use of only 11 concepts and 210 images (yielding 2310 observations per language) is load-bearing for the generalizability of the 'structural penalty' claim; without broader concept coverage or cross-dataset validation, it remains possible that the observed IoU gaps (+0.114 base, +0.143 large) are specific to the chosen visual stimuli rather than a language-wide text-branch property.
[Results] Results on scaling (Basque Δ=-0.056, Luxembourgish Δ=-0.076, Arabic Δ=+0.033): the separation of corpus-coverage versus tokeniser-fertility failures is central to the second finding, yet the manuscript provides no quantitative breakdown (e.g., fertility statistics or pretraining token overlap per language) to substantiate that the widening gaps are not confounded by metric sensitivity or image-specific factors.

minor comments (2)

[Abstract/Methods] Abstract and methods: the exact procedure for computing the English reference in Spearman rank correlation and cluster-mask IoU should be stated more explicitly to allow full reproduction of the cross-language agreement scores.
[Discussion] The energy-consumption claim (3.4-3.9 Wh per 1000 queries) is a useful practical contribution but lacks a short derivation or reference to the measurement protocol in the main text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive evaluation and recommendation of minor revision. We address the major comments point by point below, with planned revisions where appropriate.

read point-by-point responses

Referee: [Evaluation] Evaluation section: the use of only 11 concepts and 210 images (yielding 2310 observations per language) is load-bearing for the generalizability of the 'structural penalty' claim; without broader concept coverage or cross-dataset validation, it remains possible that the observed IoU gaps (+0.114 base, +0.143 large) are specific to the chosen visual stimuli rather than a language-wide text-branch property.

Authors: The evaluation employs a compact set of 11 concepts and 210 images to support dense, multi-metric analysis with 2,310 paired observations per language, enabling high-powered statistical tests (Wilcoxon p < 10^{-300}) while holding the visual encoder fixed. This design isolates text-branch effects more rigorously than broader but less controlled evaluations. The structural penalties appear consistently across cluster-mask IoU, top-percentile IoU, and Spearman correlation at both scales, reducing the chance of stimulus-specific artifacts. We agree that broader coverage would strengthen generalizability claims and will revise the manuscript to add an explicit limitations discussion on evaluation scope along with suggestions for future cross-dataset validation. revision: yes
Referee: [Results] Results on scaling (Basque Δ=-0.056, Luxembourgish Δ=-0.076, Arabic Δ=+0.033): the separation of corpus-coverage versus tokeniser-fertility failures is central to the second finding, yet the manuscript provides no quantitative breakdown (e.g., fertility statistics or pretraining token overlap per language) to substantiate that the widening gaps are not confounded by metric sensitivity or image-specific factors.

Authors: The scaling results show opposing trends—widening gaps for Basque and Luxembourgish versus improvement for Arabic—which we interpret as evidence for distinct failure modes (tokenizer fertility versus corpus coverage). Although the current manuscript does not report explicit fertility rates or token-overlap statistics, the patterns align with known XLM-RoBERTa pretraining imbalances. To address potential confounds, we will incorporate quantitative tokenizer analysis (fertility statistics and subword overlap with English) for the affected languages into the revised results section. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper contains no derivations, fitted parameters, or self-referential definitions. All claims rest on direct empirical measurements: the visual encoder (ViT-B/32 or ViT-H/14) is held fixed while only the XLM-RoBERTa text branch receives language-specific inputs; performance gaps are quantified via cluster-mask IoU, top-percentile IoU, and Spearman correlation computed against an external English reference on 2310 paired observations per language. Statistical tests (Wilcoxon) and cross-scale consistency are reported without any reduction of outputs to inputs by construction. No load-bearing self-citations or uniqueness theorems appear.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract describes an empirical probing study using standard CLIP and XLM-RoBERTa components; no free parameters, axioms, or invented entities are introduced or fitted.

pith-pipeline@v0.9.0 · 5648 in / 1279 out tokens · 81256 ms · 2026-05-12T02:12:28.077026+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

dense multilingual CLIP probe in which the visual encoder is held identical across thirteen typologically diverse languages and only the XLM-RoBERTa text branch varies

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages · 5 internal anchors

[1]

Learning Transferable Visual Models From Natural Language Supervision

A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” inProceedings of the 38th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 139. PMLR, 2021, pp....

work page internal anchor Pith review Pith/arXiv arXiv 2021
[2]

LLaV A-NeXT: Improved reasoning, OCR, and world knowledge,

H. Liu, C. Li, Y . Li, B. Li, Y . Zhang, S. Shen, and Y . J. Lee, “LLaV A-NeXT: Improved reasoning, OCR, and world knowledge,” January 2024, blog post. [Online]. Available: https://llava-vl.github.io/ blog/2024-01-30-llava-next/

work page 2024
[3]

How Far Are We to GPT-4V? Closing the Gap to Commercial Multimodal Models with Open-Source Suites

Z. Chen, W. Wang, H. Tian, S. Ye, Z. Gao, E. Cui, W. Tong, K. Hu, J. Luo, Z. Ma, J. Wang, R. Dong, L. Ding, W. Su, X. Zhu, L. Lu, B. Li, T. Lu, Y . Qiao, and J. Dai, “How far are we to GPT-4V? Closing the gap to commercial multimodal models with open-source suites,”arXiv preprint arXiv:2404.16821, 2024. [Online]. Available: https://arxiv.org/abs/2404.16821

work page internal anchor Pith review arXiv 2024
[4]

Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

P. Wang, S. Bai, S. Tan, S. Wang, Z. Fan, J. Bai, K. Chen, X. Liu, J. Wang, W. Ge, Y . Fan, K. Dang, M. Du, X. Ren, R. Men, D. Liu, C. Zhou, J. Zhou, and J. Lin, “Qwen2-VL: Enhancing vision-language model’s perception of the world at any resolution,”arXiv preprint arXiv:2409.12191, 2024. [Online]. Available: https://arxiv.org/abs/2409.12191

work page internal anchor Pith review Pith/arXiv arXiv 2024
[5]

Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone

M. Abdin, J. Aneja, H. Awadalla, A. Awasthi, A. A. Awan, N. Bach, A. Bahree, A. Bakhtiari, J. Bao, H. Behlet al., “Phi-3 technical report: A highly capable language model locally on your phone,” 2024. [Online]. Available: https://arxiv.org/abs/2404.14219

work page internal anchor Pith review Pith/arXiv arXiv 2024
[6]

XTREME: A massively multilingual multi-task benchmark for evaluating cross-lingual generalisation,

J. Hu, S. Ruder, A. Siddhant, G. Neubig, O. Firat, and M. John- son, “XTREME: A massively multilingual multi-task benchmark for evaluating cross-lingual generalisation,” inProceedings of the 37th International Conference on Machine Learning. PMLR, 2020, pp. 4411–4421

work page 2020
[7]

Un- supervised cross-lingual representation learning at scale,

A. Conneau, K. Khandelwal, N. Goyal, V . Chaudhary, G. Wenzek, F. Guzm ´an, E. Grave, M. Ott, L. Zettlemoyer, and V . Stoyanov, “Un- supervised cross-lingual representation learning at scale,”Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 8440–8451, 2020

work page 2020
[8]

MEGA: Multilingual evaluation of generative AI,

K. Ahuja, H. Diddee, R. Hada, M. Ochieng, K. Ramesh, P. Jain, A. Nambi, T. Ganu, S. Segal, M. Ahmedet al., “MEGA: Multilingual evaluation of generative AI,”Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp. 4232–4267, 2023

work page 2023
[9]

Chatgpt beyond english: Towards a comprehensive evaluation of large language models in multilingual learning,

V . D. Lai, N. T. Ngo, A. P. B. Veyseh, H. Man, F. Dernoncourt, T. Bui, and T. H. Nguyen, “Chatgpt beyond english: Towards a comprehensive evaluation of large language models in multilingual learning,”Findings of the Association for Computational Linguistics: EMNLP 2023, pp. 13 171–13 189, 2023

work page 2023
[10]

How good is your tokenizer? on the monolingual performance of multilingual language models,

P. Rust, J. Pfeiffer, I. Vuli ´c, S. Ruder, and I. Gurevych, “How good is your tokenizer? on the monolingual performance of multilingual language models,” inProceedings of the 59th Annual Meeting of the Association for Computational Linguistics, 2021, pp. 3118–3135

work page 2021
[11]

Language model tokenizers introduce unfairness between languages,

A. Petrov, E. La Malfa, P. Torr, and A. Bibi, “Language model tokenizers introduce unfairness between languages,”Advances in Neural Information Processing Systems, vol. 36, 2024

work page 2024
[12]

Sentence-BERT: Sentence embeddings using Siamese BERT-networks,

N. Reimers and I. Gurevych, “Sentence-BERT: Sentence embeddings using Siamese BERT-networks,” inProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, 2019, pp. 3982– 3992

work page 2019
[13]

Cross-platform evaluation of reasoning capabilities in foun- dation models,

J. de Curt `o, I. de Zarz `a, P. Garc ´ıa, J. Cabot, J. C. Cano, and C. T. Calafate, “Cross-platform evaluation of reasoning capabilities in foun- dation models,”Information Processing & Management, vol. 63, no. 7, Part B, p. 104878, 2026

work page 2026
[14]

Energy-aware multilingual vision–language models for drone smart sensing,

J. de Curt `o, M. Liz, I. de Zarz `a, and C. T. Calafate, “Energy-aware multilingual vision–language models for drone smart sensing,”Drones, vol. 10, no. 5, 2026. [Online]. Available: https://www.mdpi.com/ 2504-446X/10/5/361

work page 2026
[15]

arxiv preprint arXiv:1805.04687 (Apr 2020)

F. Yu, H. Chen, X. Wang, W. Xian, Y . Chen, F. Liu, V . Madhavan, and T. Darrell, “BDD100K: A diverse driving dataset for heterogeneous multitask learning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2020, pp. 2636–2645. [Online]. Available: https://arxiv.org/abs/1805.04687

work page arXiv 2020
[16]

AI Energy Score: Standardized energy efficiency ratings for AI models,

Hugging Face, S. Luccioni, Y . Jernite, R. Pierrard, I. Moutawwakil, M. Mitchell, B. Gamazaychikov, S. Chamberlin, S. Hooker, C.-J. Wu, and E. Strubell, “AI Energy Score: Standardized energy efficiency ratings for AI models,” 2025, accessed: February 2026. [Online]. Available: https://huggingface.github.io/AIEnergyScore/

work page 2025
[17]

Power hungry processing: Watts driving the cost of AI deployment?

A. S. Luccioni, S. Viguier, and A.-L. Ligozat, “Power hungry processing: Watts driving the cost of AI deployment?”Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, pp. 85–99, 2024

work page 2024
[18]

XGLUE: A new benchmark dataset for cross-lingual pre-training, understanding and generation,

Y . Liang, N. Duan, Y . Gong, N. Wu, F. Guo, W. Qi, M. Gong, L. Shou, D. Jiang, G. Caoet al., “XGLUE: A new benchmark dataset for cross-lingual pre-training, understanding and generation,” inProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, 2020, pp. 6008–6018

work page 2020
[19]

Comparative analysis of reasoning capabilities in foundation models,

J. de Curt `o and I. de Zarz `a, “Comparative analysis of reasoning capabilities in foundation models,” in2024 2nd International Conference on Foundation and Large Language Models (FLLM), 2024, pp. 141–149

work page 2024
[20]

Do llamas work in English? on the latent language of multilingual transformers,

C. Wendler, V . Veselovsky, G. Monea, and R. West, “Do llamas work in English? on the latent language of multilingual transformers,” inProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 2024, pp. 15 366–15 394

work page 2024
[21]

Extract free dense labels from CLIP,

C. Zhou, C. C. Loy, and B. Dai, “Extract free dense labels from CLIP,” inEuropean Conference on Computer Vision (ECCV), 2022, pp. 696– 712

work page 2022
[22]

One map to find them all: Real-time open-vocabulary mapping for zero-shot multi-object navigation,

F. L. Busch, T. Homberger, J. Ortega-Peimbert, Q. Yang, and O. Ander- sson, “One map to find them all: Real-time open-vocabulary mapping for zero-shot multi-object navigation,” inIEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 14 835–14 842

work page 2025
[23]

OpenCLIP,

G. Ilharco, M. Wortsman, R. Wightman, C. Gordon, N. Carlini, R. Taori, A. Dave, V . Shankar, H. Namkoong, J. Miller, H. Hajishirzi, A. Farhadi, and L. Schmidt, “OpenCLIP,” Zenodo, 2021

work page 2021
[24]

LAION-5B: An open large-scale dataset for training next generation image–text models,

C. Schuhmann, R. Beaumont, R. Vencu, C. Gordon, R. Wightman, M. Cherti, T. Coombes, A. Katta, C. Mullis, M. Wortsmanet al., “LAION-5B: An open large-scale dataset for training next generation image–text models,” inAdvances in Neural Information Processing Systems (NeurIPS), 2022

work page 2022
[25]

Energy and policy consider- ations for deep learning in nlp,

E. Strubell, A. Ganesh, and A. McCallum, “Energy and policy consider- ations for deep learning in nlp,”Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp. 3645–3650, 2019

work page 2019
[26]

Green ai,

R. Schwartz, J. Dodge, N. A. Smith, and O. Etzioni, “Green ai,” Communications of the ACM, vol. 63, no. 12, pp. 54–63, 2020

work page 2020
[27]

Carbon Emissions and Large Neural Network Training

D. Patterson, J. Gonzalez, Q. Le, C. Liang, L.-M. Munguia, D. Rothchild, D. So, M. Texier, and J. Dean, “Carbon emissions and large neural network training,”arXiv preprint arXiv:2104.10350, 2021

work page internal anchor Pith review Pith/arXiv arXiv 2021
[28]

Towards the systematic reporting of the energy and carbon footprints of machine learning,

P. Henderson, J. Hu, J. Romoff, E. Brunskill, D. Jurafsky, and J. Pineau, “Towards the systematic reporting of the energy and carbon footprints of machine learning,”Journal of Machine Learning Research, vol. 21, no. 248, pp. 1–43, 2020

work page 2020
[29]

Measuring the carbon intensity of AI in cloud instances,

J. Dodge, T. Prewitt, R. T. des Combes, E. Odber, R. Schwartz, E. Strubell, A. S. Luccioni, N. A. Smith, N. DeCario, and W. Buchanan, “Measuring the carbon intensity of AI in cloud instances,”Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Trans- parency, pp. 1877–1894, 2022

work page 2022
[30]

The use of ranks to avoid the assumption of normality implicit in the analysis of variance,

M. Friedman, “The use of ranks to avoid the assumption of normality implicit in the analysis of variance,”Journal of the American Statistical Association, vol. 32, no. 200, pp. 675–701, 1937

work page 1937
[31]

Llm-powered cooperative perception framework for mixed uav-vehicle platoons,

J. de Curt `o and I. de Zarz `a, “Llm-powered cooperative perception framework for mixed uav-vehicle platoons,” in2025 10th International Conference on Fog and Mobile Edge Computing (FMEC), 2025, pp. 282–289

work page 2025