Pith. sign in

REVIEW 3 major objections 26 references

A DRL policy trained only on small carbon-aware scheduling instances transfers to harder ones and beats static tuning once problems diverge.

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 ↗

DRL dynamic algorithm configuration trained on small carbon-aware flow-shop instances generalizes and outperforms static tuning as instance complexity grows.

T0 review reviewed 2026-07-13 challenge →

load-bearing objection The attached full text is a completely different paper (T2I jailbreaks); the DRL-DAC carbon-aware scheduling claims cannot be audited at all. the 3 major comments →

arxiv 2604.01886 v2 pith:65NADLQ7 submitted 2026-04-02 math.OC cs.NE

When does learning pay off? A study on DRL-based dynamic algorithm configuration for carbon-aware scheduling

Andrea Mencaroni , Robbert Reijnen , Yingqian Zhang , Dieter Claeys This is my paper
classification math.OC cs.NE MSC 90B3568T0590C59
keywords dynamic algorithm configurationdeep reinforcement learningcarbon-aware schedulingpermutation flow-shopparameter controltransfer learningevolutionary algorithms
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.

The reading

The paper asks when the high cost of training a deep reinforcement learning controller for dynamic algorithm configuration is justified. The authors train such a controller exclusively on small, simple carbon-aware permutation flow-shop instances, then freeze it and deploy it on both similar and more complex unseen instances. Against a statically tuned baseline prepared under comparable conditions, the learned policy performs about as well on easy instances that resemble training, but pulls ahead steadily as instance characteristics diverge and computational difficulty rises. The result is that the training investment pays off precisely when static tuning cannot adapt to changing problem scenarios. In short, generalization across instance types, not raw performance on the training distribution, is what makes learning worthwhile.

Core claim

Deep reinforcement learning can acquire a dynamic algorithm-configuration policy from small, simple carbon-aware flow-shop instances that remains effective on unseen instances of different types and scales; the policy matches a static baseline on cheap, similar instances and continuously outperforms it as complexity and distributional distance increase.

What carries the argument

The DRL-based Dynamic Algorithm Configuration (DAC) policy: an online parameter-control policy for an evolutionary algorithm, trained only on small instances and then transferred without further training.

Load-bearing premise

That a single static configuration, tuned under the same computational budget and on the same small-instance family used for DRL training, is a fair and strong enough baseline so that later gains on hard instances can be attributed to dynamic control rather than to under-tuning of the static competitor.

What would settle it

Retune the static baseline separately on each complex instance family under a matched total wall-clock budget; if the retuned static method then matches or beats the transferred DRL policy on those hard instances, the claim that learning pays off via generalization fails.

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

If this is right

  • A single DRL controller trained once on cheap instances can replace repeated static retuning when the instance distribution shifts.
  • Static parameter tuning of evolutionary algorithms is insufficient for carbon-aware scheduling under changing problem scales.
  • Transfer performance of DAC policies becomes a practical criterion for deciding whether the training investment is justified.
  • The value of dynamic control grows with the computational complexity and distributional distance of the target instances.

Where Pith is reading between the lines

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

  • The same transfer pattern may appear in other online configuration settings that face carbon or energy constraints, such as packing or vehicle routing.
  • A practical break-even calculator could estimate how far a target instance distribution must sit from the training set before DRL training cost is recovered.
  • Hybrid runtimes that keep a static configuration for easy instances and invoke the DRL policy only on hard ones could further reduce total compute.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 0 minor

Summary. The submission is identified as arXiv:2604.01886, a study claiming that a DRL-based Dynamic Algorithm Configuration (DAC) policy for carbon-aware permutation flow-shop scheduling, trained only on small/simple instances, transfers to unseen and more complex instances and increasingly outperforms a statically tuned baseline as instance characteristics diverge—thereby justifying the training cost. The abstract alone states this transfer and outperformance narrative. The body supplied as the full manuscript, however, is an unrelated paper on low-effort prompt-based jailbreak attacks against text-to-image safety filters (taxonomy of artistic reframing, material substitution, etc., with ASR tables on SDv1.5 and other T2I models). No methods, instance generators, DRL architecture, static-tuning protocol, budgets, tables, or results for the carbon-aware scheduling study appear in the document.

Significance. If the abstract’s claims were substantiated—robust transfer of a DRL-DAC policy trained only on small carbon-aware PFSP instances, with clear gains over a fair static baseline precisely when instances become harder—the result would be practically and methodologically useful for deciding when expensive online learning of algorithm control is warranted in real-world scheduling. That significance cannot be assessed: the load-bearing experimental design, baseline fairness, and quantitative results are absent from the supplied manuscript. The T2I jailbreak body that is present is a separate contribution and does not support the claimed scheduling results.

major comments (3)
  1. Manuscript identity mismatch: title, paper_id (2604.01886), and abstract describe a DRL-DAC study on carbon-aware PFSP; the full text is the unrelated T2I jailbreak paper (Low-Effort Jailbreak Attacks Against Text-to-Image Safety Filters). No § on problem formulation, DRL-DAC framework, training distribution, transfer protocol, or static baseline exists for the claimed work. The central claim that learning “pays off” under transfer cannot be audited from the document as submitted.
  2. Fairness of the static baseline (the reader’s weakest assumption) is load-bearing for attributing gains to dynamic control rather than under-tuning or unequal wall-clock budgets. The abstract asserts a “fair point of comparison,” but the manuscript contains no tuning budget, search space, instance family for tuning, or matched-compute protocol. Without those, the transfer/outperformance claim is unsupported.
  3. Transfer and scaling claims (“as instance characteristics diverge and computational complexities increase, the DRL-learned policy continuously outperforms static tuning”) require held-out instance generators, complexity axes, statistical tests, and error bars. None of these appear; only the abstract asserts them. The abstract alone is insufficient to establish that the training investment is worthwhile.

Circularity Check

0 steps flagged

No circularity can be audited: supplied full text is a different paper (T2I jailbreaks), so the DRL-DAC scheduling claims have no derivation chain present to reduce.

full rationale

The abstract for arXiv:2604.01886 describes an empirical DRL-based Dynamic Algorithm Configuration study on carbon-aware permutation flow-shop scheduling: train a policy only on small simple instances, transfer to unseen harder instances, and compare against a statically tuned baseline. That design is ordinary experimental transfer evaluation, not a first-principles derivation that could be circular by construction. However, the CACHEABLE PAPER SOURCE CONTEXT and FULL MANUSCRIPT TEXT provided are the unrelated paper “Low-Effort Jailbreak Attacks Against Text-to-Image Safety Filters” (arXiv:2604.01888). No equations, instance generators, training protocol, static-tuning budget, tables, or results for the scheduling claim appear. With no load-bearing derivation, fit-as-prediction, or self-citation chain present for 2604.01886, there is nothing that reduces by construction to its inputs. Score 0; steps empty. Circularity of the actual scheduling paper cannot be assessed from the supplied text.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

Abstract-only review of an empirical optimization paper. Load-bearing premises are domain modeling choices and experimental fairness assumptions rather than free physical constants or invented particles. No formal axioms or new entities are introduced in the abstract; free parameters would be DRL and EA hyperparameters fitted during training/tuning but are not numerically reported here.

free parameters (3)
  • DRL training hyperparameters and network architecture
    Policy performance and transfer depend on learning rates, network size, reward shaping, and episode design chosen during training on small instances; values not given in the abstract.
  • Static baseline tuning budget and search space
    Fairness of the comparison hinges on how thoroughly the static configuration was tuned; that budget is a free experimental choice that can reverse the 'learning pays off' conclusion.
  • Instance generator parameters for 'small/simple' vs 'complex'
    The claimed generalization depends on how training and test instance distributions are defined (job counts, machine counts, carbon intensity profiles); those knobs are not specified in the abstract.
axioms (3)
  • domain assumption A static tuned configuration of the evolutionary algorithm is a fair and competitive baseline for judging whether dynamic DRL control is worth the training cost.
    Central comparative claim in the abstract rests on this baseline being adequately tuned under comparable resources.
  • domain assumption Policies learned only on small, simple carbon-aware PFSP instances can be deployed without retraining on larger or distributionally shifted instances.
    The transfer experiment design assumes zero-shot deployment of the trained controller is meaningful.
  • domain assumption Carbon-aware permutation flow-shop scheduling is a representative real-world setting in which the value of DAC generalization can be assessed.
    The paper's broader claim that 'learning pays off' under complexity is scoped through this problem class.

reviewed 2026-07-13 · how reviews work

0 comments
Cite this review

Pith. "Pith review of When does learning pay off? A study on DRL-based dynamic algorithm configuration for carbon-aware scheduling." pith.science (2026). https://pith.science/paper/65NADLQ7

@misc{pith2026260401886,
  author       = {Pith},
  title        = {Pith review of: When does learning pay off? A study on DRL-based dynamic algorithm configuration for carbon-aware scheduling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/65NADLQ7}},
  note         = {Machine review of arXiv:2604.01886}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Deep reinforcement learning (DRL) has recently emerged as a promising tool for Dynamic Algorithm Configuration (DAC), enabling evolutionary algorithms to adapt their parameters online rather than relying on static tuned configurations. While DRL can learn effective control policies, training is computationally expensive. This cost may be justified if learned policies generalize, allowing the training effort to transfer across instance types and problem scales. Yet, for real-world optimization problems, it remains unclear whether this promise holds in practice and under which conditions the investment in learning pays off. In this work, we investigate this question in the context of the carbon-aware permutation flow-shop scheduling problem. We develop a DRL-based DAC framework and train it exclusively on small, simple instances. We then deploy the learned policy on both similar and more complex unseen instances and compare its performance against a static tuned baseline, which provides a fair point of comparison. Our findings show that the proposed method provides a strong dynamic algorithm control policy that can be effectively transferred to different unseen problem instances. Notably, on simple and cheap to compute instances, similar to those observed during training and tuning, DRL performs comparably with the statically tuned baseline. However, as instance characteristics diverge and computational complexities increase, the DRL-learned policy continuously outperforms static tuning. These results confirm that DRL can acquire robust and generalizable control policies which are effective beyond the training instance distributions. This ability to generalize across instance types makes the initial computational investment worthwhile, particularly in settings where static tuning struggles to adapt to changing problem scenarios.

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

26 extracted references · 7 linked inside Pith

  1. [1]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923, 2025. 4

    Pith/arXiv arXiv 2025

  2. [2]

    Ai auditing: The bro- ken bus on the road to ai accountability

    Abeba Birhane, Ryan Steed, Victor Ojewale, Briana Vec- chione, and Inioluwa Deborah Raji. Ai auditing: The bro- ken bus on the road to ai accountability. In2024 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML), pages 612–643. IEEE, 2024. 3

    2024

  3. [3]

    Jailbreaking text-to-image models with llm- based agents.arXiv preprint arXiv:2408.00523, 2024

    Yingkai Dong, Zheng Li, Xiangtao Meng, Ning Yu, and Shanqing Guo. Jailbreaking text-to-image models with llm- based agents.arXiv preprint arXiv:2408.00523, 2024. 2

    Pith/arXiv arXiv 2024

  4. [4]

    Hotflip: White-box adversarial examples for text classifica- tion

    Javid Ebrahimi, Anyi Rao, Daniel Lowd, and Dejing Dou. Hotflip: White-box adversarial examples for text classifica- tion. InProceedings of the 56th Annual Meeting of the Asso- ciation for Computational Linguistics (Volume 2: Short Pa- pers), pages 31–36, 2018. 3

    2018

  5. [5]

    Colornet: Investigat- ing the importance of color spaces for image classification

    Shreyank N Gowda and Chun Yuan. Colornet: Investigat- ing the importance of color spaces for image classification. InAsian conference on computer vision, pages 581–596. Springer, 2018. 1

    2018

  6. [6]

    Stegcolnet: Steganal- ysis based on an ensemble colorspace approach

    Shreyank N Gowda and Chun Yuan. Stegcolnet: Steganal- ysis based on an ensemble colorspace approach. InJoint IAPR International Workshops on Statistical Techniques in Pattern Recognition (SPR) and Structural and Syntactic Pat- tern Recognition (SSPR), pages 313–323. Springer, 2021. 3

    2021

  7. [7]

    Saidakhror Gulyamov, Said Gulyamov, Andrey Rodionov, Rustam Khursanov, Kambariddin Mekhmonov, Djakhongir Babaev, and Akmaljon Rakhimjonov. Prompt injection at- tacks in large language models and ai agent systems: A com- prehensive review of vulnerabilities, attack vectors, and de- fense mechanisms.Information, 17(1):54, 2026. 3

    2026

  8. [8]

    Evilpromptfuzzer: generating inappropriate con- tent based on text-to-image models.Cybersecurity, 7(1):70,

    Juntao He, Haoran Dai, Runqi Sui, Xuejing Yuan, Dun Liu, Hao Feng, Xinyue Liu, Wenchuan Yang, Baojiang Cui, and Kedan Li. Evilpromptfuzzer: generating inappropriate con- tent based on text-to-image models.Cybersecurity, 7(1):70,

  9. [9]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 1

    2016

  10. [10]

    Perception-guided jailbreak against text-to-image models

    Yihao Huang, Le Liang, Tianlin Li, Xiaojun Jia, Run Wang, Weikai Miao, Geguang Pu, and Yang Liu. Perception-guided jailbreak against text-to-image models. InProceedings of the AAAI Conference on Artificial Intelligence, pages 26238– 26247, 2025. 2, 3

    2025

  11. [11]

    A watermark for large language models

    John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, and Tom Goldstein. A watermark for large language models. InInternational conference on machine learning, pages 17061–17084. PMLR, 2023. 3

    2023

  12. [12]

    A survey of attacks on large vision- language models: Resources, advances, and future trends

    Daizong Liu, Mingyu Yang, Xiaoye Qu, Pan Zhou, Yu Cheng, and Wei Hu. A survey of attacks on large vision- language models: Resources, advances, and future trends. arXiv preprint arXiv:2407.07403, 2024. 1

    Pith/arXiv arXiv 2024

  13. [13]

    Treant: Red-teaming text-to-image models with tree-based semantic transformations

    Yi Liu, Guowei Yang, Gelei Deng, Feiyue Chen, Yuqi Chen, Ling Shi, Tianwei Zhang, and Yang Liu. Treant: Red-teaming text-to-image models with tree-based semantic transformations. 2

  14. [14]

    Sora: A review on background, technology, limitations, and opportunities of large vision models.arXiv preprint arXiv:2402.17177, 2024

    Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jian- feng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models.arXiv preprint arXiv:2402.17177, 2024. 4

    Pith/arXiv arXiv 2024

  15. [15]

    Pla: Prompt learning attack against text-to-image generative models

    Xinqi Lyu, Yihao Liu, Yanjie Li, and Bin Xiao. Pla: Prompt learning attack against text-to-image generative models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16851–16860, 2025. 4, 5

    2025

  16. [16]

    Jailbreaking prompt attack: A controllable adversarial attack against diffusion models

    Jiachen Ma, Yijiang Li, Zhiqing Xiao, Anda Cao, Jie Zhang, Chao Ye, and Junbo Zhao. Jailbreaking prompt attack: A controllable adversarial attack against diffusion models. arXiv preprint arXiv:2404.02928, 2024. 2

    Pith/arXiv arXiv 2024

  17. [17]

    Coljailbreak: Collaborative generation and editing for jailbreaking text-to-image deep generation.Advances in Neural Information Processing Systems, 37:60335–60358,

    Yizhuo Ma, Shanmin Pang, Qi Guo, Tianyu Wei, and Qing Guo. Coljailbreak: Collaborative generation and editing for jailbreaking text-to-image deep generation.Advances in Neural Information Processing Systems, 37:60335–60358,

  18. [18]

    Midjourney v5.2, 2023

    Midjourney, Inc. Midjourney v5.2, 2023. AI-generated im- age. 1

    2023

  19. [19]

    An overview of image steganography

    Tayana Morkel, Jan HP Eloff, and Martin S Olivier. An overview of image steganography. InIssa, pages 1–11, 2005. 3

    2005

  20. [20]

    Anyone can jailbreak: Prompt-based attacks on llms and t2is.arXiv preprint arXiv:2507.21820, 2025

    Ahmed B Mustafa, Zihan Ye, Yang Lu, Michael P Pound, and Shreyank N Gowda. Anyone can jailbreak: Prompt-based attacks on llms and t2is.arXiv preprint arXiv:2507.21820, 2025. 1

    Pith/arXiv arXiv 2025

  21. [21]

    Say it differently: Linguistic styles as jailbreak vectors.arXiv preprint arXiv:2511.10519,

    Srikant Panda and Avinash Rai. Say it differently: Linguistic styles as jailbreak vectors.arXiv preprint arXiv:2511.10519,

    arXiv

  22. [22]

    Red teaming language models with language models

    Ethan Perez, Saffron Huang, Francis Song, Trevor Cai, Ro- man Ring, John Aslanides, Amelia Glaese, Nat McAleese, and Geoffrey Irving. Red teaming language models with language models. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3419–3448, 2022. 1

    2022

  23. [23]

    Unsafe diffusion: On the generation of unsafe images and hateful memes from text- to-image models

    Yiting Qu, Xinyue Shen, Xinlei He, Michael Backes, Sav- vas Zannettou, and Yang Zhang. Unsafe diffusion: On the generation of unsafe images and hateful memes from text- to-image models. InProceedings of the 2023 ACM SIGSAC conference on computer and communications security, pages 3403–3417, 2023. 5

    2023

  24. [24]

    Hierarchical text-conditional image gener- ation with clip latents.arXiv preprint arXiv:2204.06125, 1 (2):3, 2022

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener- ation with clip latents.arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2

    Pith/arXiv arXiv 2022

  25. [25]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1, 2, 3, 4, 5

    2022

  26. [26]

    Photorealistic text-to-image diffusion models with deep

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep

This paper was first reviewed by grok-4.5 on July 13, 2026.