- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies prospective recovery decisions after neural training damage. Its central premise is that two training states can be indistinguishable under current task behavior yet have different optimal recovery actions because their optimizer states differ. The authors formalize this as recovery-specific observation aliasing and introduce the Self-Dynamics Network (SDN) framework, in which a scorer estimates the future cost associated with each admissible recovery action before committing to one. Candidate recovery branches are executed during development to generate action-specific supervisory targets; at deployment, the scorer operates on the information available at decision time rather than observing those future outcomes.
The paper builds its argument through several controlled experiments. First, it constructs matched-state pairs using momentum SGD on a sine-regression problem and AdamW on a digits-classification problem. In the AdamW experiment, one clone receives an optimizer-state update using true labels and another receives an update using 50%-corrupted labels, after which the parameters are restored exactly. The resulting states therefore have matched parameters and represented functions but different optimizer histories. Across the reported confirmatory pairs, 33/80 sine pairs and 79/80 AdamW pairs prefer different retain/reset actions. Second, the authors study optimizer non-equivariance under function-preserving ReLU rescaling, contrasting Adam with a symmetry-compatible Path-SGD control. They also examine whether admissible internal interventions can affect recovery.
The main policy experiments evaluate state-aware action selection on held-out AdamW cohorts. In the binary retain/reset task, a loss-only reference has a mean cost of 64.6614, compared with 50.0788 for the reported classical state-aware reference and 50.2302 for the neural SDNN scorer. In the four-action task, both the classical and neural state-conditioned policies have lower reported cost than the hindsight best fixed action on the final cohort. A separate seven-action Transformer teacher evaluates recovery selection using explicit 500-step future rollouts before committing. The results are promising in the controlled tasks, but the Transformer experiment is not operationally equivalent to the state-only SDN policy and incurs substantially greater update-equivalent compute.
The paper also reports useful negative evidence. Its classical and neural scorers perform almost identically; a matched-state test in another setting fails to reproduce the reversal pattern; fresh sine-controller experiments do not detect statistically significant neural-versus-fixed-controller gains; and the larger-model experiments do not execute competing recovery branches. The authors accordingly acknowledge that natural prevalence, large-scale competing recovery, and operational cost benefits remain unresolved.
Critical strengths
- A well-motivated problem with a concrete counterfactual construction
The most compelling contribution is the explicit separation between current predictive behavior and the complete training state needed to determine future optimization. This is a meaningful distinction for recovery: identical current predictions do not guarantee identical future trajectories after the same intervention. The matched-state construction is considerably stronger than a correlation study. By restoring parameters after the optimizer-state write, the authors control the current parameterization and represented function while manipulating optimizer history. The use of paired action branches then makes it possible to compare recovery outcomes from the same decision-time state. This provides a credible proof of possibility that restricted task-behavior observations can fail to identify the preferred recovery action.
- Appropriate emphasis on action value rather than future prediction alone
The paper correctly distinguishes predicting a future quantity from choosing an action that improves a recovery objective. Its self-prediction experiments show that high correlation with future dynamics is not, by itself, evidence that a model will select a better intervention. This distinction is conceptually useful. A recovery controller should ultimately be evaluated by the consequences of its decisions, not only by its ability to predict future loss or recovery time. The held-out binary and four-action experiments move beyond the diagnostic prediction tasks and test that more consequential objective.
- Controlled investigation of optimizer non-equivariance
The comparison between Adam and a symmetry-compatible Path-SGD control is a useful mechanistic component. The reported function preservation, nonzero Adam equivariance defect, and subsequent path divergence connect a specific optimizer property to different future training trajectories. The control helps distinguish optimizer-induced divergence from a simple failure to preserve the represented function at the intervention boundary. The authors appropriately limit this conclusion to one controlled mechanism rather than claiming that optimizer non-equivariance explains every recovery failure.
- Relatively careful reporting of limitations and negative results
The paper is commendably explicit about the scope of its evidence. It distinguishes state-aware decision experiments from larger-scale outcome-prediction diagnostics, identifies the Transformer teacher's extra information and compute requirements, and acknowledges that the main experiments do not establish natural prevalence or large-scale operational benefit. The absence of a detected neural-function-class advantage is also reported rather than obscured. The near-parity between classical and neural scorers is informative: it suggests that the useful information may lie in the decision-time representation and action-value formulation, rather than in the use of a neural scorer specifically.
Critical weaknesses
1: The Adam moment-reset action is confounded with bias-correction behavior
This is the most important methodological concern.
The manuscript specifies that RESET clears Adam's first- and second-moment buffers while preserving the optimizer step counter. However, Adam uses that counter to compute bias-correction factors. Clearing the moments without resetting or otherwise correcting the counter therefore changes the effective scale of the next adaptive parameter update.
To see why, let the first post-reset gradient be , with both moment buffers initially zero. The moment recurrences produce
v_{t+1}=(1-\beta_2)g^2.$$ If the old step counter is retained, the bias-corrected terms are $$\widehat m_{t+1} =\frac{(1-\beta_1)g}{1-\beta_1^{t+1}}, \qquad \widehat v_{t+1} =\frac{(1-\beta_2)g^2}{1-\beta_2^{t+1}}.$$ For a large counter and negligible numerical $\epsilon$, the magnitude of the adaptive gradient term approaches $$\left|\frac{\widehat m_{t+1}} {\sqrt{\widehat v_{t+1}}}\right| \approx \frac{1-\beta_1}{\sqrt{1-\beta_2}}.$$ With standard Adam settings $\beta_1=0.9$ and $\beta_2=0.999$, this factor is approximately $3.16$. By contrast, resetting the counter as well as the moment buffers gives the usual first-step bias correction, for which the corresponding ratio is approximately $1$, ignoring $\epsilon$. Thus the adaptive gradient component of the first post-reset update can differ substantially depending on whether the counter is retained. This calculation does not prove that the reported recovery improvements are caused by the discrepancy, but it shows that the RESET intervention changes more than the availability of accumulated moments. This distinction matters because the paper interprets the results as evidence for the prospective value of optimizer history. The experiment certainly evaluates a well-defined intervention, but it does not cleanly isolate moment-history removal from the accompanying bias-correction effect. The difference could influence the relative costs of KEEP and RESET and, consequently, the action-value labels used to train the recovery policy. Required clarification or experiment: Report the exact AdamW hyperparameters and counter semantics. Compare at least the existing moment-reset/counter-retained action with a moment-and-counter reset, and ideally include a control that matches the effective first post-reset update scale. If the current action is intentionally defined as moment reset with clock retention, the claims and terminology should consistently reflect that specific intervention. This concern does not invalidate the generic observation-aliasing result, but it materially weakens the interpretation of the AdamW results as isolating optimizer-moment history. Critical weakness 2: The evidence for a general recovery framework is substantially narrower than the framing suggests The headline 79/80 AdamW action reversals come from a deliberately constructed optimizer-state perturbation: one update is performed using corrupted labels, followed by exact restoration of the parameters. This design is valuable for isolating optimizer-state effects, but it is not equivalent to demonstrating that naturally occurring neural training damage frequently produces the same type of recovery-action ambiguity. The authors acknowledge that natural prevalence is unknown. The problem is that the paper's broader claims concern state-aware prospective recovery, while the strongest action-selection evidence comes from controlled, relatively small systems. The larger-scale measurements are mostly diagnostic: the manuscript explicitly states that competing recovery branches were not executed at those scales. Consequently, the large-model experiments do not establish that the proposed policy chooses better interventions for an actual large-model training incident. There is also meaningful negative evidence within the paper. The MNIST matched-state test reports 0/30 sine-direction reversals and an interaction with the opposite sign. The fresh sine-controller experiments do not detect significant neural-versus-fixed-controller gains. These results do not negate the existence of a carefully engineered counterexample, but they indicate that the effect and policy benefit should not be presumed to transfer across settings. The paper therefore supports a narrower conclusion than a general recovery framework: decision-time state can be useful in specific controlled recovery environments. It does not yet establish how broadly this advantage generalizes, how often such ambiguity occurs in natural training failures, or whether state-conditioned selection is reliable for realistic high-cost training runs. Suggested improvement: More sharply distinguish the proof-of-possibility contribution from the efficacy of a broadly applicable recovery framework. An independent replication under a different damage mechanism, optimizer-reset definition, or model/task setting would materially strengthen the empirical contribution. If that evidence is unavailable, the claims should remain explicitly confined to the tested decision problems. 3: The contribution of state-aware information is not completely separated from the contribution of the scorer and training procedure The central algorithmic question is whether additional optimizer- and gradient-state information improves recovery decisions. The paper compares a loss-only reference against state-aware classical and neural scorers, and it acknowledges in the continuation/checkpoint experiment that the estimator and representation change together. The comparisons would be easier to interpret if the restricted and state-aware observations were evaluated using identical model classes, hyperparameter-selection budgets, training partitions, and decision rules. In particular, the same logistic or histogram-gradient-boosting estimator should be trained once on restricted behavior features and once on the extended state-aware features. The neural scorer should receive an analogous ablation. Without this matched comparison, the performance difference does not fully isolate the value of additional state information from differences in model selection or fitting. The near-identical performance of classical and neural scorers is helpful, but it does not by itself resolve the information-versus-estimator question. This is not a demand for neural methods to outperform classical ones. Rather, it is a request to make the causal claim about the value of state-aware information as clean as the paper's matched-state intervention. 4: The formal results are sound at a basic level but provide limited new theory Proposition 1 formalizes the excess cost incurred when two states with the same restricted observation require different actions. Its two-action expression is a useful characterization of the decision problem, but it is a direct consequence of the definition of action value and the standard logic of partial-observation decision theory. Proposition 2 similarly formalizes the consequence of optimizer equivariance and bounds trajectory divergence by accumulating one-step defects under Lipschitz conditions. The argument is reasonable, but the conclusion follows through a standard recurrence bound once the state transformation and update equivariance assumptions are established. The paper appropriately acknowledges that the generic aliasing principle and optimizer-state dependence are established ideas. Even so, the theoretical sections do not provide a new guarantee for learning the recovery policy, estimating action values, or generalizing to previously unseen recovery states. Most of the novel contribution is empirical and conceptual rather than a new theorem or algorithmic guarantee. The related-work boundary also deserves a more explicit treatment. For example, Guo (2026), Delayed Optimizer-State Transport Shapes Short-Horizon Training Decisions directly studies optimizer-state transport in prospective training-intervention decisions, while Or (2026), Automatic Stability and Recovery for Neural Network Training studies automatic recovery after training instability. The manuscript cites these works, so the issue is not simply missing citations. It is that the precise incremental contribution of SDN relative to existing optimizer-aware intervention selection and training-recovery systems could be more sharply demonstrated. 5: The Transformer teacher does not establish deployment-equivalent recovery performance The seven-action Transformer experiment uses explicit 500-step shadow rollouts for the candidate actions before selecting one. The manuscript correctly identifies this as a different operating regime from a state-only SDN policy. The reported teacher and identity comparison also does not reach significance on the binary recovered/unrecovered outcome under the stated McNemar test ((p=0.0703)); the comparison with the predeclared fixed gauge is significant ((p=0.0215)). The continuous (V_{50}) difference and binary recovery test measure different outcomes, so the results are not inherently contradictory. Nevertheless, the evidence of improvement depends on the endpoint and comparator. Moreover, the teacher requires 5,000 recovery-update equivalents versus 2,000 for one identity trajectory, and wall-clock time, energy, and I/O are not measured. This experiment is appropriately described as a teacher or rollout-enabled extension, but it should not be interpreted as direct evidence that the state-only SDN policy is useful at Transformer scale. The distinction is especially important because the larger-model diagnostics do not evaluate competing recovery actions. Additional reproducibility concerns The paper provides a useful split summary, including fit, validation, and final cohorts, and states that final cohorts are held out from later tuning. However, the provided PDF does not expose all details necessary to reproduce the complete intervention and scorer pipeline. Important details include the precise AdamW settings, reset-clock implementation, damage-generation protocol, exact cost and utility definitions for every experiment, the hyperparameter-selection procedure for each scorer, and the unit used to calculate uncertainty in the different cohorts. The authors mention a source package and preserved experimental records. Providing executable intervention-generation and evaluation scripts, together with the complete scorer-selection configuration, would make the claims much easier to independently verify.Questions for authors
- How sensitive are the AdamW findings to the reset-clock convention?
The paper resets both moment buffers while retaining the optimizer step counter. Can the authors report results for a reset that also resets the counter, alongside the current intervention? Please report the exact values and whether the action reversals and held-out policy gains persist after controlling for the first post-reset effective update scale. This answer would materially affect my confidence in the main AdamW interpretation.
- Can the value of the state-aware information be isolated using matched scorer ablations?
Can the authors evaluate loss-only and state-aware observations using the same logistic-regression and histogram-gradient-boosting pipelines, with matching cohort splits, tuning budgets, and selection criteria? An analogous ablation for SDNN would help establish whether the observed gain is genuinely attributable to additional decision-time information rather than differences in model fitting.
- What exactly is supported by the larger-model experiments?
The paper reports large-model damage/outcome diagnostics but states that competing recovery branches were not executed at those scales. Could the authors identify the largest setting where the actual action-conditioned recovery policy was trained and evaluated, and clearly separate that result from outcome prediction and observational recovery logs? A bounded additional experiment in which at least two candidate recovery actions are compared under matched conditions would materially strengthen the scope of the contribution.
- How should the negative cross-setting result affect the stated claim?
The paper reports 0/30 matched-state reversals in its MNIST test, with an opposite interaction sign, while the fresh sine-controller comparisons do not detect a significant gain. What property of the intervention or training protocol accounts for the discrepancy with the confirmatory AdamW result? Do the authors have evidence that predicts when the aliasing construction will be effective rather than treating transfer as an open question?
- Limitations and Ethical Considerations
The paper is transparent about several major limitations: natural prevalence is unknown, the larger-scale studies do not execute competing recovery actions, the neural scorer shows no clear advantage over the classical scorers, and operational costs have not been measured. These are appropriate disclosures.
The remaining issue is the distinction between the demonstrated proof of possibility and the implied scope of a prospective recovery framework. The paper should ensure that the abstract and conclusion preserve this boundary. The reset-clock confound also deserves explicit treatment in the limitations or main experimental discussion, since it affects how the primary optimizer-reset results should be interpreted.
I see no clear ethics or Code of Ethics concern in the supplied manuscript. The experiments do not appear to involve human subjects or sensitive personal data, and the stated application is training stability and recovery. I would not flag the paper for ethics review based on the evidence available.
- Additional / Minor Suggestions
Define every recovery cost and utility quantity in one consistent location, including the normalization used for the cumulative-loss metrics and the precise interpretation of each recovery endpoint.
Clearly distinguish the state-only SDN policy, the classical state-aware scorers, the rollout-enabled Transformer teacher, and any retrospective oracle. They have different information and compute budgets and should not be conflated.
Provide an experiment-level table listing the independent evaluation unit, the number of independent states or seeds, the paired structure, the uncertainty method, and the relevant holdout boundary.
Report the action-selection agreement with the future-aware minimum together with the realized cost. Agreement alone is insufficient when action costs are heterogeneous, and cost alone can conceal whether the controller reliably chooses the correct action.
Temper claims about the Transformer extension to reflect the nonsignificant binary comparison against identity, the dependence on the chosen recovery metric, and the extra rollout computation.
Clarify the intended role of the two propositions. The formalism is useful, but the paper would be stronger if it explicitly stated which claims are established by the elementary decision-theoretic results and which require the new experiments.
2 comments
Response to zgN9 (2/2): theory, negative findings, compute, and reproducibility
-
Theory and incremental novelty. The earlier observation-aliasing proposition and optimizer-equivariance recurrence are useful formalizations of established reasoning; the revision presents different, precisely scoped results. Section 2 derives a conditional Omega_P(N log N) first-hitting-time lower bound for a specified near-collinear reference gradient flow, necessary signed-curvature and physical-energy conditions for the four-Gaussian objective, and an exact residual-contraction identity for a local three-knot hat coefficient correction. Existing ReLU/spline literature supplies related tools and precedents. The full coupled SDNN algorithm in Section 3 is distinct from the reference flow, and the paper claims no general convergence theorem for it.
-
Former Transformer experiment. The seven-action teacher's candidate rollouts and extra update-equivalent cost gave it a different information and compute budget from the state-only policy. The original recovery-rate comparison to identity had p=0.0703, with recovery speed measured separately. These claims are not used to establish the revised SDNN results. In the new Table 4, external baselines receive equal observations but have substantially lower realized parameter energy than SDNN; this precludes a resource-matched superiority conclusion.
-
Negative evidence. The original MNIST matched-state probe produced 0/30 sine-direction reversals with an opposite interaction sign, and the fresh sine-controller experiments lacked significant neural-versus-fixed gains. These limited transfer of the original recovery results. The new study transparently reports its own limitations: 0/24 successes at 0.75d spacing and historical long-run forgetting examples (Tables 5–6). These are distinct tests, not a purported replication of the old probes.
-
Reproducibility and endpoints. The prepared reproducibility archive contains the revision's LaTeX, fixed learner source, experimental scripts, seed-level records, and the matched-state diagnostic analyses. We distinguish frozen outputs from independently rerun checks and do not equate observed worker time with matched FLOPs. The revised endpoint is four-population squared-error risk, evaluated for old and new populations; the original recovery cost and action utility are different endpoints. Accordingly, the current claims are restricted to the explicitly specified learner, synthetic observation stream, and achieved-resource regime.
Thank you for insisting that mechanism, evidence scope, and causal attribution be presented with mathematical precision.
Response to zgN9 (1/2): AdamW accounting and the continuity of the mechanism
We appreciate the reviewer's detailed technical examination and the recognition of the original paper's distinction between current function and future training dynamics. We address the important original confound precisely and explain the mathematical connection to the current revision.
AdamW moment reset and step counter. The reviewer's calculation identifies a genuine compound intervention. When moment buffers are zeroed but a large step counter is retained, the first post-reset adaptive ratio (for nonzero gradient and negligible epsilon) approaches (1-beta_1)/sqrt(1-beta_2), approximately 3.162 for beta_1=0.9 and beta_2=0.999. The reported original comparison therefore concerns moment reset with counter retention; attributing it solely to erasing momentum history would require the separate counter-reset and update-scale controls requested. The revised SDNN evidence uses an independently specified ReLU learning rule and does not rely on the old 79/80 AdamW reversal statistic.
An explicit connection between the two investigations. The original Section 3 and Table 2 analyze function-preserving ReLU rescalings and their optimizer-dependent future trajectories. The new Section 2 studies the physical Jacobian and the local correction directions; Section 3 uses slice knot/amplitude coordinates invariant under positive ReLU rescaling and a balanced physical lift. Thus both address the gap between equivalent current functions and their underlying learning-state geometry, although the mechanisms and evaluation objectives differ. This relation is algebraically explicit, while the old action values and the new learning-risk endpoints remain separate mathematical quantities.
Designed incidents and scope. The revision reports 150 confirmation runs under its specified four-Gaussian weight-damage and stream-switch protocol, together with 40 matched-state transport-rule experiments. It also records 0/24 successes when population spacing is tightened to 0.75d, and historical long continuations with forgetting (Section 4, Tables 5–6). The original large-model measurements were diagnostics without competing branches; no large-model recovery-selection inference is carried into the new manuscript.
Information-versus-estimator identification. A matched restricted/full-observation classifier experiment would have sharpened the earlier action-selection inference. The new experiment asks a different identification question: after a common state at observation 560, how do specified transport rules change subsequent function learning? It holds the supervised/replay/cap protocol fixed and reports seed-level endpoints, while later branch trajectories evolve independently (Section 4, Figure 4).