pith. sign in

arxiv: 2511.07969 · v2 · submitted 2025-11-11 · 💻 cs.CL

Unified Work Embeddings: Contrastive Learning of a Bidirectional Multi-task Ranker

Matthias De Lange , Jens-Joris Decorte , Jeroen Van Hautte This is my paper

Pith reviewed 2026-05-18 00:00 UTC · model grok-4.3

classification 💻 cs.CL
keywords unified work embeddingscontrastive learningmulti-task rankinglabor market intelligencezero-shot generalizationbi-encoderworkbench benchmarkinfonce loss
0
0 comments X

The pith

A single bi-encoder trained on shared work tasks ranks entirely new labor-market targets without fine-tuning.

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

The paper sets out to establish that work-related data contains enough shared structure for one compact model to learn useful representations across many ranking tasks at once. If this holds, it would replace the current practice of building separate task-specific systems for problems like skill matching or job title search, each with its own training data and latency overhead. The authors introduce WorkBench, a benchmark that casts six real-world work tasks as ranking problems drawn from ontologies and annotations, then train a task-agnostic bi-encoder with a many-to-many InfoNCE loss and soft late interaction on token embeddings. A sympathetic reader would care because the resulting model shows zero-shot performance on unseen target spaces while running with two orders of magnitude fewer parameters than the strongest generalist baseline and a 4.4 MAP gain.

Core claim

Unified Work Embeddings (UWE) is a task-agnostic bi-encoder that exploits the structure of work-related data through a many-to-many InfoNCE objective and task-agnostic soft late interaction, delivering zero-shot ranking on unseen target spaces in the work domain together with low-latency inference using far fewer parameters than generalist models such as Qwen3-8B and a 4.4 MAP improvement.

What carries the argument

The many-to-many InfoNCE objective combined with soft late interaction on token-level embeddings inside a bidirectional bi-encoder.

If this is right

  • Joint training on the six tasks in WorkBench produces positive cross-task transfer.
  • The same embeddings support ranking on entirely new target spaces inside the work domain without retraining.
  • Inference remains low-latency because the model uses two orders of magnitude fewer parameters than the best generalist alternatives.
  • A single model can replace multiple specialized systems while improving accuracy on labor-market benchmarks.

Where Pith is reading between the lines

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

  • The same joint-contrastive approach could be tested on other domains that share consistent textual structure, such as medical coding or regulatory text.
  • Adding explicit handling of job-description length or skill hierarchies might further strengthen zero-shot transfer to new ontologies.
  • Production systems could measure end-to-end latency and cost savings when replacing multiple fine-tuned models with one UWE instance.

Load-bearing premise

The structure shared across work-related tasks is rich enough that one contrastive training run produces representations that transfer to completely new target spaces without any task-specific fine-tuning or extra supervision.

What would settle it

Curate one additional ranking task from a fresh work-related ontology never seen in training and measure whether UWE still outperforms both task-specific baselines and large generalist models on that task; failure to do so would falsify the zero-shot generalization claim.

Figures

Figures reproduced from arXiv: 2511.07969 by Jens-Joris Decorte, Jeroen Van Hautte, Matthias De Lange.

Figure 1
Figure 1. Figure 1: Overview of the six WorkBench tasks, demonstrating samples [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a) Data dependency graph of the training data, showing the many-to-many and one-to-many edges from skill space [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Task transfer experiment showing our three partial [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of temperature influence for late inter [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Job title similarity threshold over the number of [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
read the original abstract

Applications in labor market intelligence demand specialized NLP systems for a wide range of tasks, characterized by extreme multi-label target spaces, strict latency constraints, and multiple text modalities such as skills and job titles. These constraints have led to isolated, task-specific developments in the field, with models and benchmarks focused on single prediction tasks. Exploiting the shared structure of work-related data, we propose a unifying framework, combining a wide range of tasks in a multi-task ranking benchmark, and a flexible architecture tackling text-driven work tasks with a single model. The benchmark, WorkBench, is the first unified evaluation suite spanning six work-related tasks formulated explicitly as ranking problems, curated from real-world ontologies and human-annotated resources. WorkBench enables cross-task analysis, where we find significant positive cross-task transfer. This insight leads to Unified Work Embeddings (UWE), a task-agnostic bi-encoder that exploits our training-data structure with a many-to-many InfoNCE objective, and leverages token-level embeddings with task-agnostic soft late interaction. UWE demonstrates zero-shot ranking performance on unseen target spaces in the work domain, and enables low-latency inference with two orders of magnitude fewer parameters than best-performing generalist models (Qwen3-8B), with +4.4 MAP improvement.

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.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces WorkBench, the first unified benchmark spanning six work-related ranking tasks (formulated from real-world ontologies and human annotations) and proposes Unified Work Embeddings (UWE), a task-agnostic bi-encoder trained with a many-to-many InfoNCE objective plus token-level soft late interaction. It reports significant positive cross-task transfer, zero-shot ranking performance on unseen target spaces within the work domain, low-latency inference, and a +4.4 MAP improvement over Qwen3-8B while using two orders of magnitude fewer parameters.

Significance. If the zero-shot generalization to entirely new target spaces holds without label leakage or semantic overlap, the work would provide a practical unifying framework for labor-market NLP, replacing multiple task-specific models with a single efficient bi-encoder and enabling cross-task analysis via the new benchmark.

major comments (2)
  1. §5 (zero-shot experiments): The claim of zero-shot ranking on unseen target spaces (novel skill/job-title spaces) is load-bearing for the central contribution, yet the evaluation does not report an explicit disjoint-label split or ablation that randomizes cross-task label alignment. Without this, observed MAP gains could arise from shared domain vocabulary and pre-existing embedding similarities rather than the many-to-many InfoNCE plus soft late interaction design.
  2. §4.1 (training objective): The many-to-many InfoNCE formulation is presented as isolating transferable work structure, but the manuscript provides no ablation that removes cross-task label alignment or measures performance when target spaces are forced to be semantically disjoint; this leaves the positive cross-task transfer result vulnerable to alternative explanations.
minor comments (2)
  1. Abstract and §5: The efficiency claim ('two orders of magnitude fewer parameters') should include the exact parameter counts for UWE versus Qwen3-8B and the latency measurements to allow direct verification.
  2. §3 (WorkBench construction): Provide more detail on how the six tasks were curated to ensure no unintended semantic overlap between training and held-out target spaces, including any ontology alignment steps.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive comments. The concerns about potential label leakage in the zero-shot setting and the need for stronger ablations on cross-task alignment are well-taken. We address each point below and will incorporate revisions to provide clearer evidence for the contributions of the many-to-many InfoNCE objective and soft late interaction.

read point-by-point responses

  1. Referee: [—] §5 (zero-shot experiments): The claim of zero-shot ranking on unseen target spaces (novel skill/job-title spaces) is load-bearing for the central contribution, yet the evaluation does not report an explicit disjoint-label split or ablation that randomizes cross-task label alignment. Without this, observed MAP gains could arise from shared domain vocabulary and pre-existing embedding similarities rather than the many-to-many InfoNCE plus soft late interaction design.

    Authors: We agree that an explicit verification of label disjointness would strengthen the zero-shot claim. WorkBench tasks are constructed from independent real-world ontologies (e.g., ESCO skills vs. O*NET occupations) with limited label intersection, as documented in Section 3; this design choice was intended to ensure unseen target spaces. To directly address the concern, we will add (i) a quantitative analysis of label overlap across tasks and (ii) an ablation that randomizes cross-task label alignments while keeping the same training data volume. These additions will appear in a revised §5 and will help isolate the role of the proposed objective and architecture from domain vocabulary effects. revision: yes

  2. Referee: [—] §4.1 (training objective): The many-to-many InfoNCE formulation is presented as isolating transferable work structure, but the manuscript provides no ablation that removes cross-task label alignment or measures performance when target spaces are forced to be semantically disjoint; this leaves the positive cross-task transfer result vulnerable to alternative explanations.

    Authors: We acknowledge that the current presentation would benefit from an ablation that explicitly removes or disrupts cross-task label alignment. The many-to-many InfoNCE is motivated by the shared structure across work-related tasks, but we agree that demonstrating robustness under forced semantic disjointness would rule out alternative explanations. In the revision we will include such an ablation in §4.1, training variants where target spaces are artificially made disjoint (via label permutation or subset selection) and comparing them to the full multi-task setup. This will clarify the contribution of cross-task alignment to the observed transfer. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical evaluation on held-out tasks is independent of model equations

full rationale

The paper trains a bi-encoder with a many-to-many InfoNCE objective on the multi-task WorkBench and reports zero-shot MAP on held-out ranking tasks with new target spaces. These performance numbers are measured directly on separate test splits rather than being algebraically forced by the training loss or parameter fits. No self-definitional loops, fitted inputs renamed as predictions, or load-bearing self-citations appear in the provided description. The cross-task transfer observation is an empirical finding on the benchmark, not a mathematical identity. The derivation chain from architecture to reported gains remains self-contained against the external benchmark splits.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the empirical effectiveness of the proposed training recipe rather than on new mathematical axioms or invented physical entities. No free parameters are explicitly introduced beyond standard contrastive-learning hyperparameters; the main assumptions are domain-specific (shared structure in labor-market text) and are tested via the benchmark rather than derived.

pith-pipeline@v0.9.0 · 5536 in / 1369 out tokens · 45054 ms · 2026-05-18T00:00:56.847620+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

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

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.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. WorkRB: A Community-Driven Evaluation Framework for AI in the Work Domain

    cs.CL 2026-03 unverdicted novelty 7.0

    WorkRB is the first open community-driven benchmark for AI in the work domain, organizing 13 tasks from 7 groups with dynamic multilingual ontology loading and modular design for proprietary task integration.

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages · cited by 1 Pith paper · 11 internal anchors

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

    work page

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

    work page

  3. [3]

    GPT-4 Technical Report

    Achiam, J.; Adler, S.; Agarwal, S.; Ahmad, L.; Akkaya, I.; Aleman, F. L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  4. [4]

    u ller, L.; T \

    Bechler-Speicher, M.; Finkelshtein, B.; Frasca, F.; M \"u ller, L.; T \"o nshoff, J.; Siraudin, A.; Zaverkin, V.; Bronstein, M. M.; Niepert, M.; Perozzi, B.; Galkin, M.; and Morris, C. 2025. Position: Graph Learning Will Lose Relevance Due To Poor Benchmarks. In Forty-second International Conference on Machine Learning Position Paper Track

    work page 2025

  5. [5]

    Bekkerman, R.; and Gavish, M. 2011. High-precision phrase-based document classification on a modern scale. In Proceedings of the 17th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '11, 231–239. New York, NY, USA: Association for Computing Machinery. ISBN 9781450308137

    work page 2011

  6. [6]

    Chen, T.; Kornblith, S.; Norouzi, M.; and Hinton, G. 2020. A simple framework for contrastive learning of visual representations. In International conference on machine learning, 1597--1607. PmLR

    work page 2020

  7. [7]

    Chen, X.; and He, K. 2021. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 15750--15758

    work page 2021

  8. [8]

    Church, K. W. 2017. Word2Vec. Natural Language Engineering, 23(1): 155--162

    work page 2017

  9. [9]

    Decorte, J.-J.; De Lange, M.; and Van Hautte, J. 2025. Multilingual JobBERT for Cross-Lingual Job Title Matching. Working Notes of the Conference and Labs of the Evaluation Forum (CLEF 2025), 4038

    work page 2025

  10. [10]

    Decorte, J.-J.; Van Hautte, J.; Deleu, J.; Develder, C.; and Demeester, T. 2022. Design of negative sampling strategies for distantly supervised skill extraction

    work page 2022

  11. [11]

    Decorte, J.-J.; Van Hautte, J.; Demeester, T.; and Develder, C. 2021 . JobBERT : understanding job titles through skills . In FEAST, ECML-PKDD 2021 Workshop, Proceedings , 9

    work page 2021

  12. [12]

    Decorte, J.-J.; Van Hautte, J.; Demeester, T.; and Develder, C. 2024. SkillMatch: Evaluating Self-supervised Learning of Skill Relatedness. arXiv preprint arXiv:2410.05006

    work page arXiv 2024

  13. [13]

    Decorte, J.-J.; Van Hautte, J.; Develder, C.; and Demeester, T. 2025. Efficient Text Encoders for Labor Market Analysis. arXiv preprint arXiv:2505.24640

    work page arXiv 2025

  14. [14]

    Decorte, J.-J.; Verlinden, S.; Van Hautte, J.; Deleu, J.; Develder, C.; and Demeester, T. 2023. Extreme multi-label skill extraction training using large language models. In AI4HR & PES 2023 : International Workshop on AI for Human Resources and Public Employment Services, Proceedings, 1--10

    work page 2023

  15. [15]

    Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), 4171--4186

    work page 2019

  16. [16]

    Gao, L.; Zhang, Y.; Han, J.; and Callan, J. 2021. Scaling deep contrastive learning batch size under memory limited setup. arXiv preprint arXiv:2101.06983

    work page arXiv 2021

  17. [17]

    Gao, T.; Yao, X.; and Chen, D. 2021. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  18. [18]

    Gasco, L.; Fabregat, H.; García-Sardiña, L.; Estrella, P.; Deniz, D.; Rodrigo, A.; and Zbib, R. 2025. Overview of the TalentCLEF 2025: Skill and Job Title Intelligence for Human Capital Management. arXiv:2507.13275

    work page arXiv 2025

  19. [19]

    Hadsell, R.; Chopra, S.; and LeCun, Y. 2006. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), volume 2, 1735--1742. IEEE

    work page 2006

  20. [20]

    He, D.; Zhao, J.; Huo, C.; Huang, Y.; Huang, Y.; and Feng, Z. 2024. A New Mechanism for Eliminating Implicit Conflict in Graph Contrastive Learning. Proceedings of the AAAI Conference on Artificial Intelligence, 38(11): 12340--12348

    work page 2024

  21. [21]

    He, K.; Fan, H.; Wu, Y.; Xie, S.; and Girshick, R. 2020. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9729--9738

    work page 2020

  22. [22]

    Javed, F.; Luo, Q.; McNair, M.; Jacob, F.; Zhao, M.; and Kang, T. S. 2015. Carotene: A Job Title Classification System for the Online Recruitment Domain. In 2015 IEEE First International Conference on Big Data Computing Service and Applications, 286--293

    work page 2015

  23. [23]

    Khattab, O.; and Zaharia, M. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, 39--48

    work page 2020

  24. [24]

    Khosla, P.; Teterwak, P.; Wang, C.; Sarna, A.; Tian, Y.; Isola, P.; Maschinot, A.; Liu, C.; and Krishnan, D. 2020. Supervised contrastive learning. Advances in neural information processing systems, 33: 18661--18673

    work page 2020

  25. [25]

    le Vrang, M.; Papantoniou, A.; Pauwels, E.; Fannes, P.; Vandensteen, D.; and De Smedt, J. 2014. Esco: Boosting job matching in europe with semantic interoperability. Computer, 47(10): 57--64

    work page 2014

  26. [26]

    Liu, Y.; Huang, L.; Giunchiglia, F.; Feng, X.; and Guan, R. 2024. Improved Graph Contrastive Learning for Short Text Classification. Proceedings of the AAAI Conference on Artificial Intelligence, 38(17): 18716--18724

    work page 2024

  27. [27]

    Loshchilov, I.; and Hutter, F. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  28. [28]

    Malandri, L.; Mercorio, F.; and Serino, A. 2025. SkiLLMo: Normalized ESCO Skill Extraction through Transformer Models. In Proceedings of the 40th ACM/SIGAPP Symposium on Applied Computing, 1969--1978

    work page 2025

  29. [29]

    Mitra, B.; and Craswell, N. 2017. Neural models for information retrieval. arXiv preprint arXiv:1705.01509

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  30. [30]

    Muennighoff, N.; Tazi, N.; Magne, L.; and Reimers, N. 2022. MTEB: Massive Text Embedding Benchmark. arXiv preprint arXiv:2210.07316

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  31. [31]

    Oord, A. v. d.; Li, Y.; and Vinyals, O. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  32. [32]

    Reimers, N.; and Gurevych, I. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  33. [33]

    Sayfullina, L.; Malmi, E.; and Kannala, J. 2018. Learning Representations for Soft Skill Matching. In van der Aalst, W. M. P.; Batagelj, V.; Glava s , G.; Ignatov, D. I.; Khachay, M.; Kuznetsov, S. O.; Koltsova, O.; Lomazova, I. A.; Loukachevitch, N.; Napoli, A.; Panchenko, A.; Pardalos, P. M.; Pelillo, M.; and Savchenko, A. V., eds., Analysis of Images, ...

    work page 2018

  34. [34]

    Song, K.; Tan, X.; Qin, T.; Lu, J.; and Liu, T.-Y. 2020. Mpnet: Masked and permuted pre-training for language understanding. Advances in neural information processing systems, 33: 16857--16867

    work page 2020

  35. [35]

    Thakur, N.; Reimers, N.; R \"u ckl \'e , A.; Srivastava, A.; and Gurevych, I. 2021. Beir: A heterogenous benchmark for zero-shot evaluation of information retrieval models. arXiv preprint arXiv:2104.08663

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  36. [36]

    Tsacoumis, S.; and Willison, S. 2010. O* NET analyst occupational skill ratings: Procedures. Alexandria, VA: Human Resources Research Organization

    work page 2010

  37. [37]

    N.; Kaiser, .; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, .; and Polosukhin, I. 2017. Attention is all you need. Advances in neural information processing systems, 30

    work page 2017

  38. [38]

    EmbeddingGemma: Powerful and Lightweight Text Representations

    Vera, H. S.; Dua, S.; Zhang, B.; Salz, D.; Mullins, R.; Panyam, S. R.; Smoot, S.; Naim, I.; Zou, J.; Chen, F.; et al. 2025. EmbeddingGemma: Powerful and Lightweight Text Representations. arXiv preprint arXiv:2509.20354

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  39. [39]

    Wang, A.; Singh, A.; Michael, J.; Hill, F.; Levy, O.; and Bowman, S. R. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  40. [40]

    Wang, L.; Yang, N.; Huang, X.; Yang, L.; Majumder, R.; and Wei, F. 2023. Improving text embeddings with large language models. arXiv preprint arXiv:2401.00368

    work page arXiv 2023

  41. [41]

    Zhang, M.; Jensen, K.; Sonniks, S.; and Plank, B. 2022. S kill S pan: Hard and Soft Skill Extraction from E nglish Job Postings. In Carpuat, M.; de Marneffe, M.-C.; and Meza Ruiz, I. V., eds., Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 4962--4984. Seattle,...

    work page 2022

  42. [42]

    Zhang, Y.; Li, M.; Long, D.; Zhang, X.; Lin, H.; Yang, B.; Xie, P.; Yang, A.; Liu, D.; Lin, J.; Huang, F.; and Zhou, J. 2025. Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models. arXiv preprint arXiv:2506.05176

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  43. [43]

    Zhu, Y.; Xu, Y.; Liu, Q.; and Wu, S. 2021. An empirical study of graph contrastive learning. arXiv preprint arXiv:2109.01116

    work page arXiv 2021

  44. [44]

    Zhu, Y.; Xu, Y.; Yu, F.; Liu, Q.; Wu, S.; and Wang, L. 2020. Deep graph contrastive representation learning. arXiv preprint arXiv:2006.04131

    work page arXiv 2020