Pith. sign in

REVIEW 4 major objections 5 minor 37 references

A single shared model trained jointly on ten heterogeneous prediction tasks from World of Tanks game state beats ten specialized single-task models on average, while using one-tenth the parameters and less training time.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 07:51 UTC pith:3BVU36D4

load-bearing objection A solid, honest applied MTL study on a large proprietary game dataset; the headline claim survives scrutiny but is under-supported by untuned STL baselines and missing confidence intervals. the 4 major comments →

arxiv 2607.21290 v1 pith:3BVU36D4 submitted 2026-07-23 cs.LG cs.AI

Multi-Task Learning for Heterogeneous Prediction from Video Game State with Transfer Learning

Jonas Pech\'e , Aliaksei Tsishurou , Alexander Zap , G\"unter Wallner This is my paper
classification cs.LG cs.AI
keywords multi-task learningvideo game analyticsgame state predictiontransfer learningloss weightinggradient balancingWorld of Tankshard parameter sharing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper asks whether one shared neural network trained jointly on ten related but heterogeneous prediction tasks—five binary classifications and five regressions—can match or beat ten separate single-task models on a dataset of 2.19 million World of Tanks battles. It reports that equal-weight multi-task training improves most tasks over single-task training, and that gradient-surgery balancing (PCGrad) gives the best aggregate trade-off across tasks, though with higher training cost. It also reports that pre-training on source tasks and then fine-tuning on target tasks helps in low-data regimes, and that pre-training on other maps improves performance on a held-out target map, with gains shrinking as target data grows. The practical stake is that one model can replace ten specialized models at one-tenth the parameters and lower inference cost, which matters for production game analytics.

Core claim

Using a shared multimodal encoder with task-specific heads, the paper finds that multi-task learning is effective for heterogeneous game-state prediction: equal-weight MTL improves the average AUC and RMSE over single-task training, and PCGrad, which projects away conflicting task gradients, achieves the best aggregate numbers among the four balancing strategies tested. The same shared representation serves as a useful initialization for fine-tuning on tasks with limited data and for adapting to a new map, with the largest relative gains at 100–1,000 training battles. The authors frame the contribution as an applied empirical study of existing MTL and transfer techniques in this setting, not

What carries the argument

Hard parameter sharing: a single encoder that fuses a rasterized mini-map image (via EfficientNet-B0), global match context, and per-unit features (via a GRU and attention) into one shared embedding, with lightweight task-specific heads for each of the ten objectives. The comparison is carried by four balancing strategies—equal weighting, random loss weighting, FAMO, and PCGrad—and by a uniform early-stopping rule on the unweighted mean validation loss. PCGrad's gradient projection is the mechanism credited with the best trade-off.

Load-bearing premise

All comparisons rely on a single early-stopping rule that monitors the unweighted mean of validation losses across tasks with very different scales, so the rankings assume this rule does not systematically favor tasks with larger loss magnitudes.

What would settle it

Re-run the RQ1/RQ2 comparisons selecting checkpoints per task (or with scale-normalized validation loss) on the same data; if equal weighting or another method overtakes PCGrad, or if single-task models overtake MTL, the paper's central ranking and the MTL claim would need revision. A second check: run the low-data fine-tuning experiments with matched total data volume between pre-trained and from-scratch pipelines to see if the transfer gains persist.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single MTL model can replace ten STL models on this task family, cutting trainable parameters by 10x and FLOPs per forward pass by 10x while improving average AUC and RMSE.
  • Gradient-aware balancing (PCGrad) yields the best aggregate metrics, but its margin over equal weighting is small and its training time is roughly 3x longer.
  • Pre-training on six source tasks improves fine-tuning on four target tasks at every data budget, with relative gains decreasing as target data grows (from about +2.4% AUC at 100 battles to +0.9% at full data).
  • Pre-training on other maps improves performance on the target map Redshire in all data regimes, most strongly at 100 battles (+20.7% AUC), which supports using existing maps to bootstrap new ones.
  • These gains depend on a single shared-model evaluation protocol; the authors caution that losses differ in scale and the early-stopping rule is 'simple and imperfect.'

Where Pith is reading between the lines

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

  • The uniform early-stopping rule averages losses with very different magnitudes (BCE around 0.5 vs MSE around 0.1), so the reported rankings could change under per-task or scale-normalized model selection; a re-analysis with task-specific checkpoints would test this.
  • The paper reports means without confidence intervals or significance tests; PCGrad's small aggregate edge over equal weighting may not be robust, and a paired significance test across seeds would clarify.
  • Because the cross-map pre-training uses far more source data than the from-scratch baseline, the gain conflates transfer with data volume; a matched-data comparison would isolate the map-diversity effect.
  • The same architecture and protocol applied to a public esports dataset with matched data budgets would show whether the findings generalize beyond this proprietary World of Tanks collection.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper studies whether multi-task learning (MTL) with hard parameter sharing improves prediction from video-game state in a large proprietary World of Tanks dataset. Ten heterogeneous tasks (binary classification and regression) are trained jointly with a shared multimodal encoder, and the authors compare single-task learning (STL) with equal-weight MTL, as well as with RLW, FAMO, and PCGrad. They further evaluate source-task pre-training followed by fine-tuning under limited target data, and cross-map transfer to a held-out map. The main claims are that equal-weight MTL improves over STL on most tasks (RQ1), that PCGrad gives the best aggregate trade-off among the weighting strategies (RQ2), that source-task pre-training improves low-data target performance (RQ3), and that pre-training on other maps improves target-map performance (RQ4).

Significance. If the central comparisons hold under more rigorous evaluation, this is a practically useful applied study: it demonstrates that one shared model can replace ten specialized models for a realistic, large-scale game-state prediction suite, with substantially lower parameter count and training cost. The paper covers a broad and relevant set of MTL optimizers and transfer setups, and the authors are unusually candid about their evaluation's limitations. However, the central empirical claims are not yet established at the level expected for a journal publication: key comparisons lack confidence intervals and significance tests, the STL baselines are not independently tuned, and the model-selection criterion is acknowledged to be imperfect. The strengths of the study are its scale (2.19M battles, 29 maps, ten tasks), the three-seed protocol, and the explicit efficiency comparison; the weaknesses are in the controls and statistical grounding.

major comments (4)
  1. [Section IV-C, Table I, Section VI] The RQ1 comparison uses the same 6.3M-parameter architecture and identical hyperparameters for STL and MTL, and the paper itself lists 'independently optimized STL baselines' as future work (Section VI). Since each STL model receives gradients from only one task while the MTL encoder receives gradients from all ten, the reported improvements could largely reflect additional training signal or implicit regularization rather than an inherent benefit of multi-task learning. To support RQ1, the authors should add per-task tuned STL baselines (e.g., learning-rate and weight-decay search, smaller backbones, or a frozen encoder with a tuned head) and report confidence intervals. Without such controls, Table I does not isolate the effect of MTL.
  2. [Table II, Section VI] No confidence intervals, error bars, or significance tests are reported for any of the quantitative comparisons. Section VI explicitly states that 'we report mean results without confidence intervals or formal significance tests.' This is particularly problematic for the RQ2 conclusion: the PCGrad versus EW AUC difference is 0.8044 versus 0.8030, a margin of 0.0014, which is almost certainly within noise. The claim that 'PCGrad produced the best overall trade-off' cannot be evaluated from the presented data. The authors should provide per-task and aggregate intervals across seeds and state whether differences are statistically meaningful.
  3. [Section IV-C] All experiments select the model checkpoint using early stopping on the unweighted mean of validation losses across tasks with very different scales (BCE around 0.5 for classification, MSE around 0.1 for regression). For MTL this selects a single joint checkpoint, while for STL the same rule reduces to per-task validation loss, so the comparison mixes model-selection protocols. The paper calls the rule 'simple and imperfect' and notes it becomes less suitable for smaller datasets. Because every RQ1–RQ4 result passes through this rule, the authors should either use scale-normalized validation losses, report sensitivity to the stopping criterion, or provide per-task early-stopping results for the shared model.
  4. [Tables III and IV, Sections V-C and V-D] The transfer experiments are not controlled for data volume. In RQ3, the scratch model sees only the target-task labels on the target subset, while the fine-tuned model has been pre-trained on all source-task labels from the full dataset. In RQ4, the scratch model sees only Redshire data while the pre-trained model has seen all other maps, which is a far larger corpus. The paper acknowledges that these are 'not controlled comparison[s] with matched total data volume.' This is an important caveat, but it also limits the strength of the causal claims in RQ3 and RQ4. The authors should add matched-data controls, e.g., pre-training on an equal-size subset of other maps (or an equal amount of source-task data), or otherwise frame the results as engineering-transfer evidence rather than evidence about transferable representations.
minor comments (5)
  1. [Section IV-C] The paper states that 'We run three seeded trials' but does not report how the three seeds are used in the tables. Please clarify whether the reported numbers are means over seeds and, if so, provide the per-seed or variance information.
  2. [Table II] The row 'Loss' should be explicitly defined: is it the final unweighted validation loss, and for which model selection? A short footnote would prevent ambiguity.
  3. [Section V-B] The statement that 'FAMO shows normalization effects on the weights, up-weighting the regression tasks but harming the classification tasks' would be more informative with a concrete example or a plot of the learned weights over training.
  4. [Section II-A] Minor language issue: 'This paper at hand focuses on mixed losses' should be revised to 'The present paper focuses on...'.
  5. [Figure 3] The subscript notation in the figure ('gmodprojg1' and 'projg2') is hard to read; please use proper subscripts and a legend that maps colors to task gradients.

Circularity Check

0 steps flagged

No significant circularity: the MTL comparison is an empirical evaluation, not a derived result.

full rationale

The paper makes no claim that reduces to its own inputs by construction. Its contribution is explicitly empirical: 'Our contribution is an applied empirical study of existing MTL and transfer-learning techniques, rather than a new MTL optimizer or architecture.' RQ1–RQ4 are answered by held-out test measurements under a fixed protocol, not by fitting a target quantity and then renaming it a prediction. The STL/MTL comparison uses 'identical architecture and optimization settings,' so the shared-encoder advantage is measured rather than assumed. The loss-weighting equations for RLW, FAMO, and PCGrad are standard and cited from external prior work; they are not derived from the paper's own results. The only self-citation, [4], supplies the architecture and input representation, but it does not provide the evidence that MTL improves generalization — that evidence comes from the paper's own experiments on a proprietary dataset. The early-stopping rule is explicitly admitted to be 'simple and imperfect,' and the limitations section notes the lack of confidence intervals and the absence of 'independently optimized STL baselines'; these are validity concerns about the empirical ranking, not circular reasoning. Likewise, the choice of Redshire as a 'representative heterogeneous target map' is a dataset-selection decision, not an equivalence between premise and conclusion. No equation in the paper is identical to another by construction, and no fitted parameter is presented as a prediction. The score reflects only the minor, non-load-bearing self-citation in the architectural starting point.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper contributes an empirical comparison, so the ledger is dominated by hand-chosen hyperparameters and modeling assumptions rather than new fitted constants or postulated entities. No new physical or architectural entity is introduced. The most load-bearing assumption is the early-stopping scalarization, acknowledged by the authors as imperfect.

free parameters (5)
  • AdamW learning rate 1e-4, weight decay 1e-3
    Hand-chosen hyperparameters held fixed across STL and MTL; not fitted, but they affect the absolute and relative results.
  • Batch size 128, embedding size 128
    Fixed architecture choices used for every model; not tuned per method.
  • FAMO inner optimizer lr 1e-3 and gamma 1e-2
    FAMO-specific hyperparameters chosen by hand (Section IV-C).
  • Early stopping patience 10 and unweighted-mean validation criterion
    Model-selection rule the authors describe as 'simple and imperfect'; it is a hand-chosen scalarization of ten task losses.
  • Low-data learning-rate reduction factor 10
    Applied to all non-full-data training to mitigate overfitting (Section V-C).
axioms (5)
  • domain assumption The ten selected prediction tasks are representative of game-state telemetry and related enough for positive transfer.
    Section III defines the tasks without an independent argument that they are related beyond 'plausible prediction targets.'
  • ad hoc to paper The unweighted mean of validation losses is a valid model-selection criterion across tasks with different scales.
    Section IV-C states the rule is 'simple and imperfect'; the entire comparison rests on it.
  • ad hoc to paper Redshire is a representative target map for cross-map transfer.
    Section V-D selects Redshire as 'representative heterogeneous target map' after the fact; no pre-registration or other-map validation.
  • domain assumption The multimodal architecture from prior work [4] (EfficientNet-B0, GRU, attention, without vision decoder) is a valid fixed backbone for comparing MTL strategies.
    Section IV-A builds on the authors' own prior model; the paper does not test whether conclusions change with a different backbone.
  • domain assumption The dataset's bias toward higher-tier, more experienced players does not invalidate relative STL-vs-MTL comparisons.
    Section III discloses the bias but assumes it affects all methods equally.

pith-pipeline@v1.3.0-alltime-deepseek · 11994 in / 9015 out tokens · 93400 ms · 2026-08-01T07:51:56.834859+00:00 · methodology

0 comments
read the original abstract

Multi-task learning (MTL) is a promising approach for prediction tasks derived from video game state data, as modern game telemetry provides multiple related supervision signals from the same structured observations. We study whether a shared model trained jointly across tasks in team-based multiplayer games can improve generalization while reducing training and inference cost compared to specialized single-task models. We adapt a multimodal architecture for endpoint prediction to a general multi-task setting that combines rasterized vision inputs, global match context, and per-unit state information through an image encoder and attention-based interaction modeling. Experiments on a large proprietary World of Tanks dataset compare single-task and multi-task training, evaluate weighting strategies for mixed losses and conflicting gradients, and test pre-training/fine-tuning under limited target-data regimes. We also examine within-game transfer across game maps under structured environment shift.

Figures

Figures reproduced from arXiv: 2607.21290 by Alexander Zap, Aliaksei Tsishurou, G\"unter Wallner, Jonas Pech\'e.

Figure 1
Figure 1. Figure 1: Left: Example World of Tanks input image on the map Sand River. Right: RGB mini-map background with additional float feature maps encoding vehicle states as Gaussian ellipsoids. evaluation protocols. Because these heads induce very differ￾ent loss scales and gradient directions, we evaluate several MTL balancing strategies (Section IV-B) for joint training. As baselines, we also report per-task single-task… view at source ↗
Figure 3
Figure 3. Figure 3: Gradient interference in MTL: task gradients can differ in both [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

37 extracted references · 1 canonical work pages

  1. [1]

    An overview of multi-task learning in deep neural networks,

    S. Ruder, “An overview of multi-task learning in deep neural networks,”

  2. [2]

    Multitask learning,

    R. Caruana, “Multitask learning,”Machine Learning, vol. 28, no. 1, pp. 41–75, 1997

    1997

  3. [3]

    Multi-task learning for dense prediction tasks: A survey,

    S. Vandenhende, S. Georgoulis, W. Van Gansbeke, M. Proesmans, D. Dai, and L. Van Gool, “Multi-task learning for dense prediction tasks: A survey,”IEEE Transactions on Pattern Analysis and Machine Intelligence, p. 1–1, 2021. [Online]. Available: http: //dx.doi.org/10.1109/TPAMI.2021.3054719

    arXiv 2021

  4. [4]

    A multimodal architecture for endpoint position prediction in team-based multiplayer games,

    J. Peche, A. Tsishurou, A. Zap, and G. Wallner, “A multimodal architecture for endpoint position prediction in team-based multiplayer games,” 2025. [Online]. Available: https://arxiv.org/abs/2507.20670

    Pith/arXiv arXiv 2025

  5. [5]

    Gra- dient surgery for multi-task learning,

    T. Yu, S. Kumar, A. Gupta, S. Levine, K. Hausman, and C. Finn, “Gra- dient surgery for multi-task learning,”arXiv preprint arXiv:2001.06782, 2020, neurIPS 2020

    Pith/arXiv arXiv 2001

  6. [6]

    Deep multi-task learning: a review of concepts, methods, and cross-domain applications,

    M. M. Abdelsamie, S. S. Azab, and H. A. Hefny, “Deep multi-task learning: a review of concepts, methods, and cross-domain applications,” International Journal of Data Science and Analytics, vol. 21, no. 1, p. 77,

  7. [7]

    Multi-task learning using uncertainty to weigh losses for scene geometry and semantics,

    A. Kendall, Y . Gal, and R. Cipolla, “Multi-task learning using uncertainty to weigh losses for scene geometry and semantics,” 2018. [Online]. Available: https://arxiv.org/abs/1705.07115

    Pith/arXiv arXiv 2018

  8. [8]

    Analytical uncertainty-based loss weighting in multi-task learning,

    L. Kirchdorfer, C. Elich, S. Kutsche, H. Stuckenschmidt, L. Schott, and J. M. K ¨ohler, “Analytical uncertainty-based loss weighting in multi-task learning,” 2024. [Online]. Available: https://arxiv.org/abs/2408.07985

    Pith/arXiv arXiv 2024

  9. [9]

    Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks,

    Z. Chen, V . Badrinarayanan, C.-Y . Lee, and A. Rabinovich, “Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks,” 2018. [Online]. Available: https://arxiv.org/abs/1711.02257

    Pith/arXiv arXiv 2018

  10. [10]

    Multiple-Gradient Descent Algorithm (MGDA),

    J.-A. D ´esid´eri, “Multiple-Gradient Descent Algorithm (MGDA),” INRIA, Research Report RR-6953, Jun. 2009, in this report, the problem of minimizing simultaneously n smooth and unconstrained criteria is considered. A descent direction common to all the criteria is identified, knowing all the gradients. An algorithm is defined in which the optimization pr...

    2009

  11. [11]

    Conflict-averse gradient descent for multi-task learning,

    B. Liu, X. Liu, X. Jin, P. Stone, and Q. Liu, “Conflict-averse gradient descent for multi-task learning,” inAdvances in Neural Information Processing Systems, 2021, arXiv:2110.14048

    Pith/arXiv arXiv 2021

  12. [12]

    Famo: Fast adaptive multitask optimization,

    B. Liu, Y . Feng, P. Stone, and Q. Liu, “Famo: Fast adaptive multitask optimization,” inAdvances in Neural Information Processing Systems, 2023, arXiv:2306.03792

    Pith/arXiv arXiv 2023

  13. [13]

    Reasonable effectiveness of random weighting: A litmus test for multi-task learning,

    B. Lin, F. Ye, Y . Zhang, and I. W. Tsang, “Reasonable effectiveness of random weighting: A litmus test for multi-task learning,” 2022. [Online]. Available: https://arxiv.org/abs/2111.10603

    Pith/arXiv arXiv 2022

  14. [14]

    Sample-level weighting for multi-task learning with auxiliary tasks,

    E. Gr ´egoire, H. Chaudhary, and S. Verboven, “Sample-level weighting for multi-task learning with auxiliary tasks,” 2023. [Online]. Available: https://arxiv.org/abs/2306.04519

    Pith/arXiv arXiv 2023

  15. [15]

    EfficientNet: rethinking model scaling for convolutional neural networks,

    M. Tan and Q. V . Le, “EfficientNet: rethinking model scaling for convolutional neural networks,” 2020. [Online]. Available: https: //arxiv.org/abs/1905.11946

    Pith/arXiv arXiv 2020

  16. [16]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” 2021. [Online]. Available: https://arxiv.org/abs/2010.11929

    Pith/arXiv arXiv 2021

  17. [17]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” 2023. [Online]. Available: https://arxiv.org/abs/1706.03762

    Pith/arXiv arXiv 2023

  18. [18]

    Grandmaster level in starcraft ii using multi-agent reinforcement learning,

    O. Vinyals, I. Babuschkin, W. M. Czarnecki, M. Mathieu, A. Dudzik, J. Chung, D. H. Choi, R. Powell, T. Ewalds, P. Georgiev, J. Oh, D. Horgan, M. Kroiss, I. Danihelka, A. Huang, L. Sifre, T. Cai, J. P. Agapiou, M. Jaderberg, A. S. Vezhnevets, R. Leblond, T. Pohlen, V . Dalibard, D. Budden, Y . Sulsky, J. Molloy, T. L. Paine, C ¸ . G¨ulc ¸ehre, Z. Wang, T. ...

    2019

  19. [19]

    Human-level control through deep reinforcement learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, and D. Hassabis, “Human-level control through deep reinforcement learning,”Nature, vol. 518, no. 7540, pp. 529–533, 2015

    2015

  20. [20]

    Impala: Scalable distributed deep-rl with importance weighted actor-learner architectures,

    L. Espeholt, H. Soyer, R. Munos, K. Simonyan, V . Mnih, T. Ward, Y . Doron, V . Firoiu, T. Harley, I. Dunning, S. Legg, and K. Kavukcuoglu, “Impala: Scalable distributed deep-rl with importance weighted actor-learner architectures,” 2018. [Online]. Available: https: //arxiv.org/abs/1802.01561

    Pith/arXiv arXiv 2018

  21. [21]

    Distral: Robust multitask reinforcement learning,

    Y . W. Teh, V . Bapst, W. M. Czarnecki, J. Quan, J. Kirkpatrick, R. Hadsell, N. Heess, and R. Pascanu, “Distral: Robust multitask reinforcement learning,” 2017. [Online]. Available: https://arxiv.org/abs/ 1707.04175

    Pith/arXiv arXiv 2017

  22. [22]

    A hybrid multi-task learning approach for optimizing deep reinforcement learning agents,

    N. Varghese and Q. Mahmoud, “A hybrid multi-task learning approach for optimizing deep reinforcement learning agents,”IEEE Access, vol. PP, pp. 1–1, 03 2021

    2021

  23. [23]

    Bootstrap your own teacher: Online policy distillation for multi- game reinforcement learning,

    D. J. Byrne, M. Tot, P. Duckworth, C. Bonnet, A. Laterre, and T. D. Bar- rett, “Bootstrap your own teacher: Online policy distillation for multi- game reinforcement learning,” in2025 IEEE Conference on Games (CoG), 2025, pp. 1–8

    2025

  24. [24]

    Decision transformer: Reinforcement learning via sequence modeling,

    L. Chen, K. Lu, A. Rajeswaran, K. Lee, A. Grover, M. Laskin, P. Abbeel, A. Srinivas, and I. Mordatch, “Decision transformer: Reinforcement learning via sequence modeling,” 2021. [Online]. Available: https://arxiv.org/abs/2106.01345

    Pith/arXiv arXiv 2021

  25. [25]

    Esta: An esports trajectory and action dataset,

    P. Xenopoulos and C. Silva, “Esta: An esports trajectory and action dataset,” 2022. [Online]. Available: https://arxiv.org/abs/2209.09861

    Pith/arXiv arXiv 2022

  26. [26]

    Knowledge enhanced graph contrastive learning for match outcome prediction,

    J. Jiang, L. Wu, Z. Hu, R. Wu, X. Shen, and H. Zhao, “Knowledge enhanced graph contrastive learning for match outcome prediction,”Information Processing & Management, vol. 62, no. 3, p. 104010, 2025. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/S0306457324003698

    2025

  27. [27]

    Virtual to real-world transfer learning: A systematic review,

    M. Ranaweera and Q. H. Mahmoud, “Virtual to real-world transfer learning: A systematic review,”Electronics, vol. 10, no. 12, 2021. [Online]. Available: https://www.mdpi.com/2079-9292/10/12/1491

    2021

  28. [28]

    Learning, attentional control, and action video games,

    C. S. Green and D. Bavelier, “Learning, attentional control, and action video games,”Current Biology, vol. 22, no. 6, pp. R197–R206, 2012

    2012

  29. [29]

    Learning general multi-agent decision model through multi-task pre-training,

    J. Wang, L. Xu, and C. Sun, “Learning general multi-agent decision model through multi-task pre-training,”Neurocomputing, vol. 627, p. 129524, 2025. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/S0925231225001961

    2025

  30. [30]

    Learning phrase representations using RNN encoder-decoder for statistical machine translation,

    K. Cho, B. van Merrienboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio, “Learning phrase representations using RNN encoder-decoder for statistical machine translation,” 2014

    2014

  31. [31]

    Pytorch image models,

    R. Wightman, “Pytorch image models,” https://github.com/rwightman/ pytorch-image-models, 2019

    2019

  32. [32]

    Swin transformer v2: Scaling up capacity and resolution,

    Z. Liu, H. Hu, Y . Lin, Z. Yao, Z. Xie, Y . Wei, J. Ning, Y . Cao, Z. Zhang, L. Dong, F. Wei, and B. Guo, “Swin transformer v2: Scaling up capacity and resolution,” 2022. [Online]. Available: https://arxiv.org/abs/2111.09883

    Pith/arXiv arXiv 2022

  33. [33]

    A convnet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” 2022. [Online]. Available: https://arxiv.org/abs/2201.03545

    Pith/arXiv arXiv 2022

  34. [34]

    Cross- stitch networks for multi-task learning,

    I. Misra, A. Shrivastava, A. Gupta, and M. Hebert, “Cross- stitch networks for multi-task learning,” 2016. [Online]. Available: https://arxiv.org/abs/1604.03539

    Pith/arXiv arXiv 2016

  35. [35]

    Modeling task relationships in multi-task learning with multi-gate mixture-of-experts,

    J. Ma, Z. Zhao, X. Yi, J. Chen, L. Hong, and E. H. Chi, “Modeling task relationships in multi-task learning with multi-gate mixture-of-experts,” inProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, ser. KDD ’18. New York, NY , USA: Association for Computing Machinery, 2018, p. 1930–1939. [Online]. Available: ...

    arXiv 2018

  36. [2017]

    Available: https://arxiv.org/abs/1706.05098

    [Online]. Available: https://arxiv.org/abs/1706.05098

    Pith/arXiv arXiv

  37. [2025]

    Available: https://doi.org/10.1007/s41060-025-00892-y

    [Online]. Available: https://doi.org/10.1007/s41060-025-00892-y

    work page doi:10.1007/s41060-025-00892-y