pith. sign in

arxiv: 2606.29983 · v1 · pith:IROKSFDEnew · submitted 2026-06-29 · 💻 cs.LG · cs.AI

Stabilizing Extrapolation in Looped Transformers via Learned Stochastic Stopping

Hsun-Yu Kuo , El Mahdi Chayti , Patrik Reizinger , Wieland Brendel , Martin Jaggi This is my paper

Pith reviewed 2026-06-30 07:26 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords looped transformersstochastic stoppinglength generalizationout-of-distribution varianceRL-Haltingalgorithmic tasksextrapolation
0
0 comments X

The pith

Stochastic loop counts during training sharply reduce OOD variance in looped transformers.

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

Looped transformers reuse the same block to handle variable-length inputs and can extrapolate beyond training lengths, yet this ability shows high variance on out-of-distribution cases. The variance traces to an unintended link between input length and the number of times the block is applied. Making the loop count random at training time breaks that link and produces stable predictions no matter how many loops are used later. A learned version of the randomization, called RL-Halting, further tunes the balance between accuracy and stability on tasks such as addition and copying.

Core claim

The paper establishes that looped transformers exhibit brittle length generalization because training creates a spurious correlation between sequence length and loop count; introducing stochasticity into the loop count at training time removes this correlation and stabilizes extrapolation, while replacing heuristic randomization with a learned stochastic schedule (RL-Halting) improves the accuracy-stability trade-off on binary addition, Dyck-1, Unique Set, and Copy.

What carries the argument

RL-Halting, a reinforcement-learning method that learns a stochastic schedule for deciding when to stop looping.

If this is right

  • Stochastic loop counts at training time reduce OOD variance and make performance consistent across different inference-time loop numbers.
  • RL-Halting learned schedules improve the accuracy-stability trade-off on binary addition, Dyck-1, Unique Set, and Copy.
  • The approach can stabilize a suboptimal computation on some tasks.
  • Treating the decision of when to stop as a training-time design choice rather than an inference-time rule aids extrapolation.

Where Pith is reading between the lines

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

  • The same stochastic-iteration idea could be tested in other iterative architectures such as recurrent networks.
  • The result points to a broader need to audit training distributions for unintended correlations between input features and computation depth.
  • Combining learned stochastic stopping with other length-generalization techniques might yield further gains on harder algorithmic problems.

Load-bearing premise

The observed out-of-distribution variance is caused mainly by the correlation between sequence length and number of loops.

What would settle it

Train models with loop counts that are fixed yet matched to length in a way that removes the correlation, then check whether high OOD variance still appears.

Figures

Figures reproduced from arXiv: 2606.29983 by El Mahdi Chayti, Hsun-Yu Kuo, Martin Jaggi, Patrik Reizinger, Wieland Brendel.

Figure 1
Figure 1. Figure 1: Oracle-over-iterations accuracy on binary addition as a function of input length for standard Transformers and Looped Transformers with fixed K = 20, length-matched K = n, and RL-Halting. RL-Halting is shown for reference and discussed in § 4.3. The shaded region marks the training-length regime (ID), and the right-hand side corresponds to longer OOD inputs. Thin solid lines denote individual runs, the bol… view at source ↗
Figure 2
Figure 2. Figure 2: Mean performance and run-to-run vari￾ability on binary addition across input lengths. Left: mean oracle-over-iterations accuracy across runs. Right: standard deviation across runs at each digit length. We compare standard Transformer base￾lines with fixed-K Looped Transformers for K ∈ {10, 15, 20}. The shaded region marks the training￾length regime (ID); longer inputs are OOD. We next vary the fixed loop b… view at source ↗
Figure 3
Figure 3. Figure 3: OOD oracle-over-iterations accuracy for stochastic variants of the length-matched schedule on binary addition. “Original” denotes deterministic K = n; stochastic variants sample K from a local window around n. Boxes show the in￾terquartile range across runs, centre lines show the median, whiskers show the min–max range, and or￾ange diamonds with error bars show the mean ± one standard deviation. We now tes… view at source ↗
Figure 4
Figure 4. Figure 4: Accuracy and prediction dynamics across inference-time loop counts on binary addition. Columns compare models trained with fixed K = 20, length-matched K = #Digits, ran￾domised length-matched K = #Digits (Random), and RL-Halting. Here K = #Digits (Random) denotes training with K = clip(n + ∆, 1, Tmax), where ∆ ∼ Unif{−5, . . . , 5}. Top row: accuracy as a function of inference-time loop count K for differe… view at source ↗
Figure 5
Figure 5. Figure 5: Toy geometry of gradient averaging. For an example z, Gk(θ; z) = ∇θLk(θ; z) is the gradient used to update the looped model param￾eters θ when training stops after k loop iterations. We write Gk = G¯ +ξk, where G¯ is the component common to gradients from nearby depths and ξk is a depth-specific residual. A single-depth update can follow one residual direction, while sampling K from nearby depths averages … view at source ↗
Figure 6
Figure 6. Figure 6: OOD variability on binary addition across models and controlled settings. Bars show the [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Policy accuracy and oracle–policy gap on Unique Set. Left: policy accuracy, evaluating each method using its prescribed stopping rule. Right: oracle-over-iterations accuracy minus policy accuracy. A large gap indicates that a correct output exists at some loop depth but is not selected by the policy. Randomised and learned stochastic schedules reduce this gap, suggesting that they stabilise predictions acr… view at source ↗
Figure 8
Figure 8. Figure 8: Learned stopping schedule and stopping entropy on binary addition. For this visualisation, the stopping distribution is truncated at Tvis = 60. Left: learned stopping distributions π (Tvis) stop,ϕ(τ = k | x) for RL-Halting, averaged over binary-addition examples with the same digit length. As the number of digits increases, the probability mass shifts toward larger loop depths, indicating that the learned … view at source ↗
read the original abstract

Looped Transformers, which repeatedly apply a shared transformer block, are an architecturally natural fit for variable-length algorithmic tasks. Although they can exhibit strong length generalization beyond the length of training sequences, this behavior is brittle, yielding high out-of-distribution (OOD) variance, even across well-performing in-distribution solutions. We trace this variance to the spurious correlation in simple algorithmic tasks between sequence length and number of loops. Introducing stochasticity into the number of loops during training sharply reduces OOD variance and stabilizes predictions across inference-time loop counts. To improve upon heuristic randomization schemes, we further analyze RL-Halting as a learned stochastic schedule and find that it generally improves the accuracy-stability trade-off. Across binary addition, Dyck-1, Unique Set, and Copy, learned stochastic stopping often improves this trade-off but can also stabilize a suboptimal computation. Our work suggests that "when to stop" should be treated as a training-time design choice, not merely an inference-time computation-allocation rule.

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 paper claims that looped transformers exhibit brittle length generalization with high OOD variance on algorithmic tasks, which it traces to a spurious correlation between training sequence length and the number of loop iterations. Introducing stochasticity over the number of loops during training reduces this variance and stabilizes predictions at inference-time loop counts; a learned RL-Halting schedule is shown to improve the accuracy-stability trade-off on binary addition, Dyck-1, Unique Set, and Copy, though it can also stabilize suboptimal computations. The work concludes that halting should be treated as a training-time design choice.

Significance. If the empirical findings hold after proper isolation of the causal factor, the result would offer a concrete training modification that improves reliability of variable-depth computation in transformers without changing the architecture, with potential applicability to other recurrent or iterative models.

major comments (2)
  1. The central attribution of OOD variance to the spurious length-loop correlation (rather than optimization dynamics or representation capacity) is load-bearing for the claim that stochasticity is causal rather than incidental. The manuscript provides no ablation that holds the loop-count distribution fixed while varying other training factors, leaving the source of variance unisolated.
  2. The statement that RL-Halting 'generally improves the accuracy-stability trade-off' but 'can also stabilize a suboptimal computation' requires quantitative support (e.g., per-task tables showing both the improvement magnitude and the cases of suboptimal stabilization) to be load-bearing for the recommendation to treat halting as a training-time choice.
minor comments (2)
  1. Notation for the stochastic loop schedule and the RL-Halting objective should be introduced with explicit equations rather than prose descriptions.
  2. The abstract and introduction would benefit from a short table summarizing the four tasks, their training lengths, and the OOD lengths used for evaluation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. Below we address each major comment point by point, indicating planned revisions to the manuscript.

read point-by-point responses

  1. Referee: The central attribution of OOD variance to the spurious length-loop correlation (rather than optimization dynamics or representation capacity) is load-bearing for the claim that stochasticity is causal rather than incidental. The manuscript provides no ablation that holds the loop-count distribution fixed while varying other training factors, leaving the source of variance unisolated.

    Authors: Our primary evidence consists of controlled comparisons between deterministic loop counts (which preserve the length-loop correlation) and stochastic loop counts (which break it), with all other training factors held fixed; the reduction in OOD variance under stochasticity supports the role of the correlation. We acknowledge, however, that this does not fully isolate the correlation from other potential contributors such as optimization dynamics. In the revision we will add an explicit discussion of this limitation and include an ablation that holds the loop-count distribution fixed while varying other training elements (e.g., optimizer settings or data ordering) where computationally feasible. revision: yes

  2. Referee: The statement that RL-Halting 'generally improves the accuracy-stability trade-off' but 'can also stabilize a suboptimal computation' requires quantitative support (e.g., per-task tables showing both the improvement magnitude and the cases of suboptimal stabilization) to be load-bearing for the recommendation to treat halting as a training-time choice.

    Authors: The manuscript already reports results across the four tasks (binary addition, Dyck-1, Unique Set, Copy) that illustrate the trade-off, including the observation that RL-Halting can stabilize suboptimal solutions. To strengthen the claim with the requested quantitative detail, the revised version will include per-task tables that explicitly list accuracy and stability metrics for each method, report the magnitude of improvements, and flag any instances of suboptimal stabilization. revision: yes

Circularity Check

0 steps flagged

No derivation chain present; purely empirical claims

full rationale

The paper reports experimental results on introducing stochastic loop counts during training of looped transformers to reduce OOD variance on algorithmic tasks. No equations, first-principles derivations, fitted parameters renamed as predictions, or self-citation chains appear in the abstract or described content. The central claims rest on observed accuracy-stability trade-offs across binary addition, Dyck-1, Unique Set, and Copy tasks, with no reduction of any result to its own inputs by construction. The work is therefore self-contained as an empirical study.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms or invented entities; the central claim rests on the empirical observation that stochasticity reduces variance.

pith-pipeline@v0.9.1-grok · 5717 in / 960 out tokens · 20960 ms · 2026-06-30T07:26:34.155538+00:00 · methodology

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

35 extracted references · 19 canonical work pages · 6 internal anchors

  1. [1]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, 9 U. V on Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, ed- itors,Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. UR...

    2017

  2. [2]

    Position: Understanding LLMs Requires More Than Statistical Generalization.arXiv,

    Patrik Reizinger, Szilvia Ujváry, Anna Mészáros, Anna Kerekes, Wieland Brendel, and Ferenc Huszár. Position: Understanding LLMs Requires More Than Statistical Generalization.arXiv,

  3. [3]

    doi: 10.48550/arxiv.2405.01964

    work page doi:10.48550/arxiv.2405.01964

  4. [4]

    Generalization on the Unseen, Logic Reasoning and Degree Curriculum, November 2024

    Emmanuel Abbe, Samy Bengio, Aryo Lotfi, and Kevin Rizk. Generalization on the Unseen, Logic Reasoning and Degree Curriculum, November 2024. URL http://arxiv.org/abs/ 2301.13105. arXiv:2301.13105 [cs]

    work page arXiv 2024

  5. [5]

    What Algorithms can Transformers Learn? A Study in Length Gen- eralization, October 2023

    Hattie Zhou, Arwen Bradley, Etai Littwin, Noam Razin, Omid Saremi, Josh Susskind, Samy Ben- gio, and Preetum Nakkiran. What Algorithms can Transformers Learn? A Study in Length Gen- eralization, October 2023. URLhttp://arxiv.org/abs/2310.16028. arXiv:2310.16028

    work page arXiv 2023

  6. [6]

    Extrapolation by Association: Length Generalization Transfer in Transformers, August 2025

    Ziyang Cai, Nayoung Lee, Avi Schwarzschild, Samet Oymak, and Dimitris Papailiopoulos. Extrapolation by Association: Length Generalization Transfer in Transformers, August 2025. URLhttp://arxiv.org/abs/2506.09251. arXiv:2506.09251 [cs]

    work page arXiv 2025

  7. [7]

    Self-Improving Transformers Overcome Easy-to-Hard and Length Generalization Challenges

    Nayoung Lee, Ziyang Cai, Avi Schwarzschild, Kangwook Lee, and Dimitris Papailiopoulos. Self-Improving Transformers Overcome Easy-to-Hard and Length Generalization Challenges. June 2025. URLhttps://openreview.net/forum?id=ZtX0MBT6mf

    2025

  8. [8]

    Universal Transformers

    Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Łukasz Kaiser. Uni- versal Transformers, July 2018. URLhttps://arxiv.org/abs/1807.03819v3

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  9. [9]

    Can You Learn an Algorithm? Generalizing from Easy to Hard Problems with Recurrent Networks, November 2021

    Avi Schwarzschild, Eitan Borgnia, Arjun Gupta, Furong Huang, Uzi Vishkin, Micah Goldblum, and Tom Goldstein. Can You Learn an Algorithm? Generalizing from Easy to Hard Problems with Recurrent Networks, November 2021. URL http://arxiv.org/abs/2106.04537. arXiv:2106.04537 [cs]

    work page arXiv 2021

  10. [10]

    End-to-end Algorithm Synthesis with Recurrent Networks: Logical Extrapolation Without Overthinking, October 2022

    Arpit Bansal, Avi Schwarzschild, Eitan Borgnia, Zeyad Emam, Furong Huang, Micah Goldblum, and Tom Goldstein. End-to-end Algorithm Synthesis with Recurrent Networks: Logical Extrapolation Without Overthinking, October 2022. URL http://arxiv.org/abs/2202. 05826. arXiv:2202.05826 [cs]

    work page arXiv 2022

  11. [11]

    Looped Transformers as Programmable Computers.arXiv, 2023

    Angeliki Giannou, Shashank Rajput, Jy-yong Sohn, Kangwook Lee, Jason D Lee, and Dimitris Papailiopoulos. Looped Transformers as Programmable Computers.arXiv, 2023. doi: 10. 48550/arxiv.2301.13196

    work page arXiv 2023

  12. [12]

    Looped Transformers for Length Generalization

    Ying Fan, Yilun Du, Kannan Ramchandran, and Kangwook Lee. Looped Transformers for Length Generalization. October 2024. URL https://openreview.net/forum?id= 2edigk8yoU

    2024

  13. [13]

    Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach

    Jonas Geiping, Sean McLeish, Neel Jain, John Kirchenbauer, Siddharth Singh, Brian R. Bar- toldson, Bhavya Kailkhura, Abhinav Bhatele, and Tom Goldstein. Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach, February 2025. URL http://arxiv.org/abs/2502.05171. arXiv:2502.05171 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  14. [14]

    Scaling Latent Reasoning via Looped Language Models

    Rui-Jie Zhu, Zixuan Wang, Kai Hua, Tianyu Zhang, Ziniu Li, Haoran Que, Boyi Wei, Zixin Wen, Fan Yin, He Xing, Lu Li, Jiajun Shi, Kaijing Ma, Shanda Li, Taylor Kergan, Andrew Smith, Xingwei Qu, Mude Hui, Bohong Wu, Qiyang Min, Hongzhi Huang, Xun Zhou, Wei Ye, Jiaheng Liu, Jian Yang, Yunfeng Shi, Chenghua Lin, Enduo Zhao, Tianle Cai, Ge Zhang, Wenhao Huang,...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  15. [15]

    Looped transformers are better at learning learning al- gorithms.arXiv preprint arXiv:2311.12424,

    Liu Yang, Kangwook Lee, Robert Nowak, and Dimitris Papailiopoulos. Looped Transformers are Better at Learning Learning Algorithms.arXiv, 2023. doi: 10.48550/arxiv.2311.12424

    work page doi:10.48550/arxiv.2311.12424 2023

  16. [16]

    PonderNet: Learning to Ponder, September

    Andrea Banino, Jan Balaguer, and Charles Blundell. PonderNet: Learning to Ponder, September

  17. [17]

    Banino, J

    URLhttp://arxiv.org/abs/2107.05407. arXiv:2107.05407 [cs]. 10

    work page arXiv

  18. [18]

    Nikunj Saunshi, Nishanth Dikkala, Zhiyuan Li, Sanjiv Kumar, and Sashank J. Reddi. Reasoning with Latent Thoughts: On the Power of Looped Transformers. October 2024. URL https: //openreview.net/forum?id=din0lGfZFd

    2024

  19. [19]

    Thinking Deeper With Recurrent Networks: Logical Extrapolation Without Overthinking

    Arpit Bansal, Avi Schwarzschild, Eitan Borgnia, Zeyad Emam, Furong Huang, Micah Goldblum, and Tom Goldstein. Thinking Deeper With Recurrent Networks: Logical Extrapolation Without Overthinking. October 2021. URLhttps://openreview.net/forum?id=kDF4Owotj5j

    2021

  20. [20]

    Simple statistical gradient-following algorithms for connectionist reinforce- ment learning

    Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforce- ment learning

  21. [21]

    Reinforcement Learning: An Overview, May 2025

    Kevin Murphy. Reinforcement Learning: An Overview, May 2025. URL http://arxiv. org/abs/2412.05265. arXiv:2412.05265 [cs]

    work page arXiv 2025

  22. [22]

    Transformers Can Achieve Length Generalization But Not Robustly, February 2024

    Yongchao Zhou, Uri Alon, Xinyun Chen, Xuezhi Wang, Rishabh Agarwal, and Denny Zhou. Transformers Can Achieve Length Generalization But Not Robustly, February 2024. URL http://arxiv.org/abs/2402.09371. arXiv:2402.09371 [cs]

    work page arXiv 2024

  23. [23]

    Randomized Positional Encodings Boost Length Generalization of Transformers, May 2023

    Anian Ruoss, Grégoire Delétang, Tim Genewein, Jordi Grau-Moya, Róbert Csordás, Mehdi Bennani, Shane Legg, and Joel Veness. Randomized Positional Encodings Boost Length Generalization of Transformers, May 2023. URL http://arxiv.org/abs/2305.16843. arXiv:2305.16843 [cs]

    work page arXiv 2023

  24. [24]

    Position Coupling: Improving Length Generaliza- tion of Arithmetic Transformers Using Task Structure

    Hanseul Cho, Jaeyoung Cha, Pranjal Awasthi, Srinadh Bhojanapalli, Anupam Gupta, and Chulhee Yun. Position Coupling: Improving Length Generaliza- tion of Arithmetic Transformers Using Task Structure. November 2024. URL https://openreview.net/forum?id=5cIRdGM1uG&referrer=%5Bthe%20profile% 20of%20Pranjal%20Awasthi%5D(%2Fprofile%3Fid%3D~Pranjal_Awasthi3)

    2024

  25. [25]

    The Impact of Positional Encoding on Length Generalization in Transformers

    Amirhossein Kazemnejad, Inkit Padhi, Karthikeyan Natesan Ramamurthy, Payel Das, and Siva Reddy. The Impact of Positional Encoding on Length Generalization in Transformers

  26. [26]

    ON V ANISHING V ARIANCE IN TRANSFORMER LENGTH GENERALIZATION

    Ruining Li and Gabrijel Boduljak. ON V ANISHING V ARIANCE IN TRANSFORMER LENGTH GENERALIZATION. 2025

    2025

  27. [27]

    How much does Initialization Affect Generalization? InProceedings of the 40th International Conference on Machine Learning, pages 28637–28655

    Sameera Ramasinghe, Lachlan Ewen Macdonald, Moshiur Farazi, Hemanth Saratchandran, and Simon Lucey. How much does Initialization Affect Generalization? InProceedings of the 40th International Conference on Machine Learning, pages 28637–28655. PMLR, July 2023. URL https://proceedings.mlr.press/v202/ramasinghe23a.html

    2023

  28. [28]

    On The Power of Curriculum Learning in Training Deep Networks

    Guy Hacohen and Daphna Weinshall. On The Power of Curriculum Learning in Training Deep Networks. InProceedings of the 36th International Conference on Machine Learn- ing, pages 2535–2544. PMLR, May 2019. URL https://proceedings.mlr.press/v97/ hacohen19a.html

    2019

  29. [29]

    Path Independent Equilibrium Models Can Better Exploit Test-Time Computation

    Cem Anil, Ashwini Pokle, Kaiqu Liang, Johannes Treutlein, Yuhuai Wu, Shaojie Bai, Zico Kolter, and Roger Grosse. Path Independent Equilibrium Models Can Better Exploit Test-Time Computation

  30. [30]

    Hierarchical Reasoning Model

    Guan Wang, Jin Li, Yuhao Sun, Xing Chen, Changling Liu, Yue Wu, Meng Lu, Sen Song, and Yasin Abbasi Yadkori. Hierarchical Reasoning Model, July 2025. URL http://arxiv.org/ abs/2506.21734. arXiv:2506.21734 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [31]

    Adaptive Computation Time for Recurrent Neural Networks

    Alex Graves. Adaptive Computation Time for Recurrent Neural Networks, February 2017. URLhttp://arxiv.org/abs/1603.08983. arXiv:1603.08983 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  32. [32]

    Dropout: A Simple Way to Prevent Neural Networks from Overfitting

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A Simple Way to Prevent Neural Networks from Overfitting

  33. [33]

    Deep Networks with Stochastic Depth

    Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Weinberger. Deep Networks with Stochastic Depth, July 2016. URL http://arxiv.org/abs/1603.09382. arXiv:1603.09382 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  34. [34]

    Reducing transformer depth on demand with structured dropout.arXiv preprint arXiv:1909.11556,

    Angela Fan, Edouard Grave, and Armand Joulin. Reducing Transformer Depth on Demand with Structured Dropout, September 2019. URL http://arxiv.org/abs/1909.11556. arXiv:1909.11556 [cs]. 11 A Implementation Details Architecture and training.Our Looped Transformer implementation follows Fan et al. [ 11]. One looped block consists of three Transformer layers a...

    work page arXiv 2019

  35. [35]

    Therefore, institutional review board approval or equivalent human-subjects review is not applicable

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...