REVIEW 4 major objections 5 minor 23 references
Not All Explanations for Deep Learning Phenomena Are Equally Valuable
T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read This paper claims that double descent, grokking, and the lottery ticket hypothesis are edge cases whose main research value is refining broad explanatory theories, not resolving them as isolated puzzles.
desk verdict A sensible, well-written position paper whose practical recommendations are useful, but whose load-bearing empirical claim about the practical irrelevance of these phenomena relies on absence-of-evidence and is weaker than it needs to be. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central conceptual machinery is the distinction between narrow ad hoc hypotheses and broad explanatory theories, paired with a sociotechnical-pragmatism principle that research value is rooted in downstream utility. The paper operationalizes the distinction with a deliberately absurd worked example: both double descent and grokking can be 'explained' post hoc by the number of prime numbers among a network's rounded parameters, a theory that tracks the effects but yields no broader insight. This example carries the argument by showing that fit to observations is not enough; an explanation must generate testable predictions or practical guidance beyond the phenomenon itself, and the paper's proposed scientific process is meant to enforce that requirement.
What would settle it
A single documented, reproducible occurrence of grokking in a large-scale production training run with standard hyperparameters, or a measured double-descent curve in a deployed model under standard regularization, would directly contradict the practical-irrelevance claim; equally, a systematic survey of practitioner logs that found no such behaviors would support it.
Extended reading notes
Core claim
The paper's central claim is that the most visible deep learning phenomena are not representative challenges from practice, and therefore the standard 'resolve the puzzle' research mode is the wrong frame. It distinguishes narrow ad hoc hypotheses—explanations overfit to the setting in which the phenomenon was observed, such as a theory that prime numbers in network parameters cause double descent and grokking—from broad explanatory theories, which are general enough to carry implications for scaling, optimization, generalization, or other core principles. The authors argue that the lasting value of double descent, grokking, and lottery tickets has already come from this second route: they forced reexamination of the bias-variance tradeoff, progress measures in training, and theories of sparsity and pruning. The normative conclusion is that researchers should prioritize utility and falsifiability when choosing which phenomena to study and which explanations to refine.
Load-bearing premise
The argument assumes that the absence of these phenomena in published scaling studies and realistic benchmark attempts means they are genuinely rare in practice, rather than merely unmeasured or unreported.
Editorial extensions
If this is right
- If the position is right, research on double descent, grokking, and lottery tickets should be judged by whether it updates a general principle, such as the bias-variance tradeoff or the meaning of progress during training, rather than by whether it matches the phenomenon's observed curves.
- Publication and funding incentives would shift toward descriptive cataloging, replication, preregistration, and shared code repositories, so that edge-case studies accumulate into a reusable body of evidence.
- Researchers would be expected to state the downstream utility of a phenomenon study up front, and explanations that are accurate but utility-free—like the prime-number parody—would be deprioritized.
- The same framework would apply to newer concerns, such as emergent abilities in large language models, where distinguishing genuinely concerning behaviors from contrived artifacts would guide whether the research is worth pursuing.
Reading between the lines
- The paper's practical-irrelevance claim could be tested by auditing production training logs and deployed-model scaling curves for signatures of double descent, grokking, or lottery-ticket substructures; if practitioners simply never measure these quantities, the claimed rarity could be a measurement gap rather than a fact about the world.
- The narrow-versus-broad distinction could be turned into a grading rubric: an explanation is broad to the extent that removing it changes a prediction about a different architecture, dataset, or training regime, and that rubric could be applied retrospectively to published explanations to see whether the paper's historical narrative holds up.
- If pragmatic prioritization were adopted, it would predict a shift in the field's output mix over time—fewer phenomenon-specific theory papers and more benchmark-style repositories and replication studies—a trend that could be checked empirically in conference proceedings.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This position paper argues that many prominent deep-learning phenomena, exemplified by double descent, grokking, and the lottery ticket hypothesis, are edge cases that rarely appear in real-world applications. On this basis it contends that resolution-oriented research—developing narrow, ad hoc explanations for each phenomenon in isolation—yields little value, and that the primary value of these phenomena lies in using them as test beds for refining broad explanatory theories of deep learning. The paper supports its position with case-by-case discussions of the three phenomena, an illustrative 'prime numbers' ad hoc explanation that is then shown to be confounded, a discussion of philosophical underpinnings, responses to alternative views, and a practical self-evaluation checklist for future research.
Significance. The paper is timely and thought-provoking, and it makes a genuinely normative proposal about research priorities rather than merely surveying phenomena. Its main strengths are its explicit scoping to edge-case phenomena, the deliberately constructed and then refuted 'prime numbers' example (Appendix A), which demonstrates awareness of confounding variables and post hoc fitting, and the practical checklist in Table 1, which is likely to be useful regardless of whether one accepts the full position. The paper is also honest about the subjectivity involved in assessing utility. If the empirical premise about practical irrelevance were established, the proposed reorientation of research effort would be significant for the community. However, the paper's central normative claim currently depends on an absence-of-evidence argument that is not fully secured.
major comments (4)
- [Section 2.1, Practical Irrelevance] The claim that double descent does not appear in real-world scaling analyses is weakened by the paper's own observation that optimal regularization removes the effect. The scaling studies cited (Kaplan et al., 2020; Hoffmann et al., 2022; Zhai et al., 2022) are run under standard, heavily regularized training regimes, so their failure to exhibit double descent may reflect a design choice rather than the phenomenon's rarity in unregularized or differently regularized practical systems. Since this absence of evidence is load-bearing for the paper's 'little value' conclusion, the authors should explicitly address this alternative explanation and state what kind of evidence would change their assessment.
- [Section 2.2, Practical Irrelevance] The grokking discussion cites a small number of unsuccessful attempts to induce grokking in realistic settings (Dziri et al., 2024) and known sensitivity to initialization scale, metric choice, and dataset size. These facts do not establish that grokking is rare in practice, because negative results are rarely reported and the cited studies were not designed as systematic searches. The text is appropriately cautious in saying attempts 'have so far been unsuccessful,' but Section 1 draws a stronger conclusion ('not a practical concern in training frontier models'). The paper should either align the strength of the conclusion with the evidence or propose a concrete, preregistered search strategy as part of its agenda.
- [Sections 2 and 3] The paper's counterfactual claim—that deliberately focusing on broad explanatory theories would be more efficient than resolution-oriented research—is not supported by the evidence presented. The 'Broader Value' paragraphs in Section 2 show that much of the broad theoretical progress attributed to these phenomena emerged from research that was initially resolution-oriented (e.g., understanding double descent through the bias-variance lens). The paper does not provide examples in which an explicit broad-theory-first approach yielded comparable practical value. Without such evidence, the recommendation to deprioritize resolution-oriented work remains an interesting conjecture rather than a well-grounded conclusion.
- [Section 1 and Section 3.1] The argument equates practical irrelevance with infrequency of appearance, but practical relevance also depends on the severity of consequences when a phenomenon does occur. Even a rare phenomenon such as grokking-like delayed generalization could impose severe costs if it occurs in a safety-critical deployment, and the paper does not weigh rarity against severity anywhere. The conclusion that 'these phenomena do not pose a significant practical challenge' does not follow from the evidence that they are seldom observed. This gap should be acknowledged and addressed, for example by restricting the claim to phenomena for which both frequency and severity are low.
minor comments (5)
- [Section 1, paragraph 2] There is a typographical error: 'what we calledge cases' should read 'what we call edge cases.'
- [Section 3 title] The heading 'On The Source of Value in Studying Deep Learning Phenomona' contains a misspelling; 'Phenomona' should be 'Phenomena.'
- [Section 2.2, last paragraph of Practical Irrelevance] The phrase 'this echos a similar point' should use 'echoes' instead of 'echos.'
- [Figure 2 caption] The grammar in 'A neural networks train set accuracy' should be 'A neural network's training set accuracy.'
- [Table 1] The self-evaluation checklist is comprehensive, but it is long relative to the body of a position paper; consider compressing the table or moving it entirely to the appendix, and cross-reference it more explicitly in Section 5.2.
Circularity Check
No significant circularity: the position paper argues from external evidence and explicitly constructs then refutes its own ad hoc explanation.
full rationale
No circular derivation was found. This is a position paper, not a predictive or fitted model: it defines a scope ('edge case phenomena') and argues normatively that such phenomena should be studied mainly as testbeds for broad explanatory theories. The practical-irrelevance claims in Section 2 are supported by external citations (e.g., Nakkiran et al., 2021; Hastie et al., 2022; Kaplan et al., 2020; Hoffmann et al., 2022; Zhai et al., 2022; Dziri et al., 2024; Miller et al., 2023; Kumar et al., 2023) rather than by fitting a parameter to a dataset and then 'predicting' a quantity that was fitted. The self-citations (Jeffares et al., 2024a; Curth et al., 2024b; Jeffares et al., 2024b) function as examples of broader value or as supporting technical analysis; they are not load-bearing premises that presuppose the paper's conclusion. The deliberately absurd 'prime numbers' explanation in Figure 4 is explicitly constructed and then dismissed as a confound in the appendix, demonstrating awareness of the exact failure mode the paper criticizes rather than committing it. The only definitional element is the scoping of 'edge case phenomena' as those without prominent practical appearance; this is a boundary for the discussion, not an empirical prediction, and the paper's normative recommendation does not reduce to this definition. The skeptical concern that practical irrelevance rests on absence-of-evidence is an evidentiary weakness, not circularity.
Assumptions & free parameters
assumptions (4)
- domain assumption Sociotechnical pragmatism, the principle that research value is rooted in downstream impact, is broadly compatible with the field's values.
- domain assumption The three selected phenomena (double descent, grokking, lottery ticket) are representative edge cases.
- domain assumption Popper's distinction between ad hoc hypotheses and falsifiable theories is applicable to machine learning explanations.
- domain assumption The absence of these phenomena in published scaling analyses or realistic experiments indicates their practical irrelevance.
Cite this review
Pith. "Pith review of Not All Explanations for Deep Learning Phenomena Are Equally Valuable." pith.science (2026). https://pith.science/paper/SB2DZT7N
@misc{pith2026250623286,
author = {Pith},
title = {Pith review of: Not All Explanations for Deep Learning Phenomena Are Equally Valuable},
year = {2026},
howpublished = {\url{https://pith.science/paper/SB2DZT7N}},
note = {Machine review of arXiv:2506.23286}
}
read the original abstract
Developing a better understanding of surprising or counterintuitive phenomena has constituted a significant portion of deep learning research in recent years. These include double descent, grokking, and the lottery ticket hypothesis -- among many others. Works in this area often develop ad hoc hypotheses attempting to explain these observed phenomena on an isolated, case-by-case basis. This position paper asserts that, in many prominent cases, there is little evidence to suggest that these phenomena appear in real-world applications and these efforts may be inefficient in driving progress in the broader field. Consequently, we argue against viewing them as isolated puzzles that require bespoke resolutions or explanations. However, despite this, we suggest that deep learning phenomena do still offer research value by providing unique settings in which we can refine our broad explanatory theories of more general deep learning principles. This position is reinforced by analyzing the research outcomes of several prominent examples of these phenomena from the recent literature. We revisit the current norms in the research community in approaching these problems and propose practical recommendations for future research, aiming to ensure that progress on deep learning phenomena is well aligned with the ultimate pragmatic goal of progress in the broader field of deep learning.
Figures
Figures from the paper (1 more)
Reference graph
Works this paper leans on
-
[7]
Heckel, R. and Yilmaz, F. F. Early stopping in deep net- works: Double descent and how to eliminate it. arXiv preprint arXiv:2007.10099,
arXiv 2007
-
[8]
Hepburn, B. and Andersen, H. Scientific Method. In Zalta, E. N. (ed.), The Stanford Encyclopedia of Philosophy . Metaphysics Research Lab, Stanford University, Summer 2021 edition,
work page 2021
-
[9]
M., Chatzimparmpas, A., Sharma, A., Watts, D
Hofman, J. M., Chatzimparmpas, A., Sharma, A., Watts, D. J., and Hullman, J. Pre-registration for predictive modeling. arXiv preprint arXiv:2311.18807,
-
[11]
B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D
Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361,
arXiv 2001
-
[12]
Kumar, T., Bordelon, B., Gershman, S
URL https://openreview.net/forum?id= 3RXAiU7sss. Kumar, T., Bordelon, B., Gershman, S. J., and Pehlevan, C. Grokking as the transition from lazy to rich training dynamics. arXiv preprint arXiv:2310.06110,
-
[14]
A survey of lottery ticket hypothesis
Liu, B., Zhang, Z., He, P., Wang, Z., Xiao, Y ., Ye, R., Zhou, Y ., Ku, W.-S., and Hui, B. A survey of lottery ticket hypothesis. arXiv preprint arXiv:2403.04861,
-
[15]
Grokking Beyond Neural Networks: An Empirical Exploration with Model Complexity
Miller, J., O’Neill, C., and Bui, T. Grokking beyond neural networks: An empirical exploration with model complex- ity. arXiv preprint arXiv:2310.17247,
-
[16]
Murty, S., Sharma, P., Andreas, J., and Manning, C. D. Grokking of hierarchical structure in vanilla transformers. arXiv preprint arXiv:2305.18741,
Show all 23 references
-
[17]
The deep boot- strap framework: Good online learners are good offline generalizers
Nakkiran, P., Neyshabur, B., and Sedghi, H. The deep boot- strap framework: Good online learners are good offline generalizers. arXiv preprint arXiv:2010.08127,
2010 arXiv
-
[21]
A., and Birdal, T
Prieto, L., Barsbey, M., Mediano, P. A., and Birdal, T. Grokking at the edge of numerical stability. arXiv preprint arXiv:2501.04697,
-
[22]
B., Ravfogel, S., and Goldberg, Y
Zaken, E. B., Ravfogel, S., and Goldberg, Y . Bitfit: Simple parameter-efficient fine-tuning for transformer- based masked language-models. arXiv preprint arXiv:2106.10199,
-
[23]
and, simultaneously, as an implicit optimization of a collective ensemble loss term (Jeffares et al., 2024b). Despite deep ensembles’ success being prevalent in practical applications (and, thus, not an edge case), approaching these theories using the principles described thro...
2024
-
[1892]
Improving reproducibility in machine learning research (a report from the neurips 2019 repro- ducibility program)
Pineau, J., Vincent-Lamarre, P., Sinha, K., Larivi `ere, V ., Beygelzimer, A., d’Alch ´e Buc, F., Fox, E., and Larochelle, H. Improving reproducibility in machine learning research (a report from the neurips 2019 repro- ducibility program). Journal of machine learning re- sear...
2019
-
[1935]
Grokking: Generalization beyond overfit- ting on small algorithmic datasets
12 Not All Explanations for Deep Learning Phenomena Are Equally Valuable Power, A., Burda, Y ., Edwards, H., Babuschkin, I., and Misra, V . Grokking: Generalization beyond overfit- ting on small algorithmic datasets. arXiv preprint arXiv:2201.02177,
-
[1988]
and Hookway, C
Legg, C. and Hookway, C. Pragmatism. In Zalta, E. N. and Nodelman, U. (eds.), The Stanford Encyclopedia of Phi- losophy. Metaphysics Research Lab, Stanford University, Winter 2024 edition,
2024
-
[2015]
Gencoglu, O., van Gils, M., Guldogan, E., Morikawa, C., S¨uzen, M., Gruber, M., Leinonen, J., and Huttunen, H
URLhttps: //web.archive.org/web/20240913143139 /https://www.cs.ox.ac.uk/people/yarin .gal/website/blog_5058.html. Gencoglu, O., van Gils, M., Guldogan, E., Morikawa, C., S¨uzen, M., Gruber, M., Leinonen, J., and Huttunen, H. Hark side of deep learning–from grad student de- sce...
1904
-
[2019]
Emergent abilities in large language models: A survey
Berti, L., Giorgi, F., and Kasneci, G. Emergent abilities in large language models: A survey. arXiv preprint arXiv:2503.05788,
-
[2020]
Forde, J. Z. and Paganini, M. The scientific method in the science of machine learning. arXiv preprint arXiv:1904.10922,
1904 arXiv
-
[2021]
Nanda, N., Chan, L., Lieberum, T., Smith, J., and Stein- hardt, J
URL https://openreview.net/forum?id= 7R7fAoUygoa. Nanda, N., Chan, L., Lieberum, T., Smith, J., and Stein- hardt, J. Progress measures for grokking via mechanistic interpretability. arXiv preprint arXiv:2301.05217,
-
[2022]
Why do random forests work? understanding tree ensembles as self-regularizing adaptive smoothers
Curth, A., Jeffares, A., and van der Schaar, M. Why do random forests work? understanding tree ensembles as self-regularizing adaptive smoothers. arXiv preprint arXiv:2402.01502, 2024a. Curth, A., Jeffares, A., and van der Schaar, M. A u-turn on double descent: Rethinking para...
-
[2023]
Deep learn- ing through a telescoping lens: A simple model provides empirical insights on grokking, gradient boosting & be- yond
Jeffares, A., Curth, A., and van der Schaar, M. Deep learn- ing through a telescoping lens: A simple model provides empirical insights on grokking, gradient boosting & be- yond. arXiv preprint arXiv:2411.00247, 2024a. Jeffares, A., Liu, T., Crabb´e, J., and van der Schaar, M. ...
-
[2024]
Sok: Pragmatic assessment of machine learning for network intrusion detection
Apruzzese, G., Laskov, P., and Schneider, J. Sok: Pragmatic assessment of machine learning for network intrusion detection. In 2023 IEEE 8th European Symposium on Security and Privacy (EuroS&P) , pp. 592–614. IEEE,
2023
-
[2025]
The values encoded in machine learning research
Birhane, A., Kalluri, P., Card, D., Agnew, W., Dotan, R., and Bao, M. The values encoded in machine learning research. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pp. 173–184,
2022
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.