acceptodds
Under review as a conference paper at ICLR 2027

Recursive Hypercompression: Phase Transitions from Self-Improvement to Self-Deterioration

Abstract

Self-improving AI systems increasingly learn from their own outputs through several feedback mechanisms at once, such as artifacts that later rounds reuse and tasks the learner selects for itself. These mechanisms are usually built and ablated one at a time, yet each can improve or deteriorate a protected objective, and all are coupled through the same learner. We show that whether such a system improves or undergoes (RSD), a decline caused by the recursion itself, is a joint property of its mechanisms, and we make that property measurable. Modeling each round as the joint evolution of the learner, an inherited defect, and the environment, we show that single-mechanism recursions do not determine the regime of the coupled one and identify what they miss: paths through both mechanisms, which a factorial ablation isolates and leave-one-out ablations cannot. When every route closes through the learner, the mechanisms share its correction budget, so coupling moves the boundary between improvement and deterioration; in linear recursions with reinforcing routes the boundary is exact, and mechanisms that each improve a quadratic protected objective alone can deteriorate it together, possibly while the system's own evaluation still reports progress. Simulations match the predicted boundaries, and in a learned self-training loop, coefficients estimated without running the combined loop predict that it amplifies although each mechanism attenuates alone. These results suggest assessing self-improving systems with factorial ablations and protected evaluation.

Primary area:
learning theory
Keywords:
recursive self-improvement, recursive self-deterioration, self-training, feedback loops, coupled dynamics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_EzPpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies recursive self-improvement, where a model learns from its own outputs (e.g. labels data that is later used for learning, evaluates its own answers, etc). It makes the argument that there are two “mechanisms” that will influence the performance of the learner. The first is what the authors call the “inherited defect” which are errors introduced by the learner in previous stages, and the “environment” which controls things like task or prompt distributions. The paper then shows that if one wants to understand whether the learner will improve in performance over time, or deteriorate, then one must consider both the “inherited defects” and the “environment” together. Treating them as independent entities may lead to misleading evaluations/conclusions.

Critical strengths
  • The math in the paper seems to be correct.

  • The experiments, while small in scale, do support the theory in the paper.

  • The main take-away of the paper is reasonable, and if it improves the way people evaluate complex AI systems then that would be a valuable contribution.

Critical weaknesses
  • While grammatically sound, the writing in this paper is difficult to follow. I think the authors were going for a high level of abstraction and terminology-focus at the cost of explanation. The result is that the "story" is lost.

  • Novelty: The main take-away of the paper is basically (paraphrased) "the whole is different from the sum of its part". This is the core challenge in multi-objective optimization and is certainly known in control theory, which the authors acknowledge since they are using the standard mathematical setups from that field. This raises a question about the novelty or significance of the work.

  • Small scale experiments: The experiments do show that the theory is useful, but they are small and serve more as a proof-of-concept. This can be fine, but the experiments were setup to illustrate the points the theory was making. What would be more useful is either larger scale experiments OR clear evidence of realistic AI deterioration in certain systems. This would help strengthen the argument that the results are significant.

Questions for authors

Do you have examples of realistic AI systems where there is deterioration caused by the interplay of "environment" and "inherited errors"?

2 comments
AuthorsOpenReview ↗

Response to Reviewer (1/2)

We thank the reviewer for the thoughtful assessment and constructive feedback.

1. Storyline and Presentation — Concern 1 (W1)

Concern: The reviewer finds the paper difficult to follow despite its grammatical clarity, noting that the high level of abstraction and emphasis on terminology obscure the central scientific storyline (W1).

Response: Thank you for highlighting the need for a clearer scientific motivation and storyline. To improve the clarity of the narrative, we have reorganized the Introduction to establish the central problem before developing the theoretical framework and its terminology.

A clearer storyline. The revised Introduction follows a more natural progression. It begins by explaining how self-improving systems learn from their own outputs and how multiple feedback mechanisms act on the same learner, despite usually being studied separately. This motivates the central question: can mechanisms that are harmless individually cause deterioration when combined, and would the system recognize such deterioration? We then introduce an everyday example, discuss what existing work establishes and leaves open, and develop the correction-budget insight before presenting the formal model and three main findings.

An intuitive example. To make the interaction concrete, we added an everyday analogy involving a person using sleeping pills and coffee to manage poor sleep. Although each remedy may be harmless alone, their interaction can create a reinforcing feedback loop that worsens sleep quality while the person feels fine. This example illustrates three central ideas: individually harmless mechanisms can fail together, their combined feedback can exceed the learner's correction budget, and internal evaluation may miss deterioration revealed by an external measure. The analogy provides an intuitive foundation for the theoretical concepts introduced later.

More accessible terminology. We also explain key concepts in plain language before introducing their mathematical formulations. For example, we introduce the correction budget as the fraction of a deviation that the learner removes by itself each round. To further improve accessibility, we added a terminology and notation table (Appendix C, Table 2) that provides definitions of the main concepts and symbols used throughout the paper.

Finally, we reorganized the discussion of prior work into a separate Related Work section (Section 2), allowing the Introduction to focus on the scientific motivation, central insight, and main contributions.

Manuscript changes

  • Section 1 (Introduction): Reorganized the narrative around the central research question, introduced an everyday analogy, and explained the correction-budget insight before presenting the formal model and main findings.
  • Section 2 (Related Work): Organized the detailed discussion of prior work separately from the Introduction to improve the clarity and flow of the narrative.
  • Appendix C (Terminology, Table 2): Added a terminology and notation table defining the main concepts and symbols used throughout the paper.

2. Novelty and Scientific Significance — Concern 2 (W2)

Concern: The reviewer questions the novelty and scientific significance of the work, arguing that the central observation that interacting mechanisms can behave differently from their individual components is already well established in classical control theory and related fields (W2).

Response: Thank you for raising the important question of how our work differs from classical control theory and existing research. We clarify our mechanism-level contributions and their relationship to classical stability theory in General Response #4.

The possibility that interacting mechanisms behave differently from isolated ones is a familiar starting point. Our contribution is to identify how feedback mechanisms interact through the learner, inherited defect, and adaptive environment, using an exact decomposition of primitive feedback paths. This reveals joint contributions that can arise even without direct coupling between the mechanisms.

In a structured model, we further characterize how the mediated effects of these mechanisms jointly determine the amplification boundary, including regimes where individually attenuating mechanisms amplify when combined or opposing feedback restores attenuation. The resulting framework connects these interactions to recursive system design parameters, providing a quantitative explanation of how different mechanisms affect the combined dynamics.

These contributions, together with their relationship to classical stability theory and representative related work, are detailed in General Response #4 (Table R7).

AuthorsOpenReview ↗

Response to Reviewer (2/2)

3. Composition-Induced Deterioration in Realistic AI Systems — Concern 3 (W3, Q1)

Concern: The reviewer notes that the experiments are relatively small-scale and primarily designed to illustrate the theory, and suggests that evidence of deterioration in realistic AI systems would strengthen the significance of the work (W3). The reviewer specifically asks for examples of realistic AI systems where deterioration is caused by the interplay between environment feedback and inherited errors (Q1).

Response: Thank you for emphasizing the need for evidence of interacting feedback mechanisms in realistic AI systems. We address this concern by identifying relevant examples from existing AI systems and conducting a new factorial experiment on RoTTA, a published continual test-time adaptation method, to directly measure the interaction between inherited-error and environment feedback. The full analysis is provided in General Response #3.

Examples from existing AI systems. Several published systems exhibit related recursive feedback mechanisms and report deterioration. In self-play reasoning, R-Zero (Huang et al., 2025) combines a challenger that generates questions (environment feedback) with a solver trained on its own majority-vote pseudo-labels (inherited errors), and reports performance degradation after several iterations. In long-horizon test-time adaptation, methods that train on self-generated targets while selecting their own data can fall below the unadapted model on long or recurring streams (PeTTA, Hoang et al., 2024; RDumb, Press et al., 2023). In self-evolving agents, Shao et al. (2025) report that accumulated memory reduces a coding agent's refusal rate on unsafe requests, although the contribution of an environment route has not been tested. These examples illustrate relevant failure modes, but do not independently isolate the contributions of the two feedback mechanisms.

Direct factorial evidence from RoTTA. To directly examine the interaction, we investigate RoTTA (Yuan et al., 2023), an existing continual test-time adaptation method whose deterioration under recurring data corruptions was independently documented by Hoang et al. (2024). Unlike our controlled simulations, this system was not designed to illustrate our theoretical framework. RoTTA contains two feedback routes: inherited errors, through teacher-generated training targets, and environment feedback, through confidence-based memory admission. We conducted a 2×22\times2 factorial experiment to isolate their individual and joint contributions. The inherited-error route alone causes substantial deterioration (+34.20 percentage points), while the environment route alone has little additional effect relative to the control. When combined, the two routes produce a +21.10 percentage-point factorial interaction (95% CI: [15.70, 26.25]; Table R6), substantially aggravating deterioration.

Interpretation and scope. The RoTTA results provide direct evidence that the interplay between inherited errors and adaptive data selection contributes to deterioration in a published AI system, beyond the individual effects of the two mechanisms. Since the inherited-error route already deteriorates alone, the experiment demonstrates composition-aggravated deterioration rather than the failure of two individually harmless mechanisms. The experiment measures the realized factorial interaction, not the spectral amplification boundary. Together with the other published examples, these findings demonstrate the relevance of interacting feedback mechanisms in realistic AI systems, while distinguishing observed deterioration from experimentally isolated interaction effects. The complete results and limitations are presented in General Response #3 (Appendix N.3, Table 20 and Figure 11).

Reviewer_nqPwOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies recursive learning systems in which a learner interacts with both a retained artifact and an adaptive environment. It decomposes perturbation propagation into paths involving neither mechanism, one mechanism, or both, and relates the joint path family to a factorial ablation under a shared background trajectory. For a minimal linear recursion with reinforcing routes, it derives an exact stability boundary in terms of mediated feedback effects sharing the learner’s correction budget. Additional results specify when amplification deteriorates a quadratic protected objective and when an insensitive reported objective delays detection. Validation includes controlled simulations, a nonlinear digit self-training loop in which separately identified mechanisms predict short-horizon amplification in the combined system, and an auxiliary test-time adaptation experiment illustrating disagreement between training progress and protected performance.

Critical strengths
  1. This paper shows a precise and interpretable account of why individually attenuating mechanisms can amplify jointly. The shared correction budget also separates the effects of consuming the learner’s available correction and strengthening a route through direct coupling. These results provide an interpretable framework for assessing mechanism composition and explain why isolated-mechanism behavior need not determine the combined system.

  2. This paper reveals that separately identified mechanisms can predict a nontrivial combined learning response. The digit experiment estimates mechanism responses without fitting a full-loop trajectory, includes a separate one-round promotion calibration, records predictions before verification. The system uses learned logistic regression, hard pseudo-labels, and a selection-based pool, so this is more informative than directly implementing the prescribed linear recurrence.

Critical weaknesses
  1. The operational connection between the path decomposition and ordinary nonlinear factorial ablations seems remaining limited. Proposition 1 is exact for derivative propagation when the retained primitive derivatives are evaluated along one common trajectory. In the nonlinear learning experiment, the four ablation arms follow different trajectories, and the paper correctly notes that their interaction is therefore not the proposition’s exact joint-family term.

  2. Although the theoretical framework connects coupled feedback mechanisms to instability and, under additional assumptions, to deterioration of the protected objective, the experiments on learned systems do not fully establish this connection. In Section 4.2, the main self-training experiment exhibits amplification consistent with the theoretical prediction, yet the protected performance continues to improve. Conversely, under stronger coupling, protected performance deteriorates while the measured amplification rate remains below one, despite a prediction above one. These observations do not necessarily contradict the local theory, but they raise concerns about whether the proposed stability boundary can reliably predict actual performance deterioration in learned systems. This limits the practical implications of the theoretical framework, particularly its usefulness as a diagnostic for recursive self-deterioration.

Questions for authors

Could the authors clarify the relationship between the predicted amplification boundary and actual protected performance deterioration in the learned self-training experiments? In particular, is there any experimental setting where predicted amplification, measured amplification, and protected performance deterioration occur consistently? If not, what explains the discrepancy between these quantities, and to what extent should the proposed boundary be interpreted as a practical predictor of recursive self-deterioration rather than local perturbation amplification?

2 comments
AuthorsOpenReview ↗

Response to Reviewer (1/2)

We thank the reviewer for the careful reading and constructive feedback.

1. Amplification and Deterioration — Concern 1 (W2, Q1)

Concern: The reviewer questions whether the theoretical connection between amplification and protected deterioration is supported by the learned-system experiments. In the original Digit experiments, measured amplification and protected performance changes appeared inconsistent, raising concerns about the practical validity of the theoretical stability boundary (W2). The reviewer asks whether predicted amplification, measured amplification, and protected deterioration agree in a concrete setting, and whether the boundary can reliably predict recursive self-deterioration rather than only local perturbation growth (Q1).

Response: Thank you for raising these important questions about the relationship between perturbation amplification and protected deterioration. We have clarified the theoretical conditions connecting them and revised the amplification-rate estimator to address the discrepancies in the Digit experiments. The revised results show close agreement between predicted and measured rates in the base configuration and between amplification and protected deterioration under stronger coupling. We summarize our findings below, with the full analysis provided in General Response #1.

Theoretical relationship (W2, Q1). Amplification concerns perturbation growth, while deterioration concerns changes in protected performance. Proposition 2 establishes their local finite-horizon relationship under protected-readout assumptions, and Proposition 5 establishes an asymptotic correspondence for the structured model under quadratic-readout and dominant-mode conditions. Thus, the spectral boundary characterizes perturbation stability, and its implication for protected deterioration depends on additional assumptions (General Response #1.1).

Main experiment (W2). With the revised estimator, the full-loop amplification rate is 0.989±0.0030.989\pm0.003, closely matching the prediction of 0.9900.990 obtained from separately identified single-mechanism responses (Table R3). Both rates are below one, consistent with protected performance continuing to improve. The original claim of full-loop amplification in this configuration has been corrected. The prediction was re-identified after the measurement rather than established prospectively.

Strong-coupling experiment (W2, Q1). At ξ=1\xi=1, the structured-model prediction is ϱ=1.013\varrho=1.013, the revised measured rate is ϱ^=1.008±0.003\hat\varrho=1.008\pm0.003, and protected performance decreases by 0.047±0.0370.047\pm0.037 nats (p<10−5p<10^{-5}; Table R4). Neither single-mechanism arm shows significant protected deterioration. This provides a concrete setting in which predicted amplification, measured amplification, and protected deterioration are consistent. The structured-model prediction additionally uses direct-coupling calibration.

Practical interpretation (Q1). The results support the amplification boundary as a useful diagnostic of local perturbation stability, with implications for protected deterioration under the stated assumptions. The learned Digit system satisfies the stationary fixed-point assumptions only approximately, and the boundary does not determine every finite-window performance trend. For example, at ξ=0.6\xi=0.6, protected performance declines while the measured rate remains just below one (0.9960.996). Predicting recursive self-deterioration therefore also requires the relevant protected-readout conditions and evidence that feedback causes the decline. These distinctions are discussed in General Response #1.

AuthorsOpenReview ↗

Response to Reviewer (2/2)

2. Path Decomposition and Factorial Ablation — Concern 2 (W1)

Concern: The reviewer questions the operational connection between theoretical path decomposition and ordinary factorial ablation in nonlinear learning systems. Proposition 1 identifies the joint-path contribution through primitive-arrow deletion along a common trajectory, whereas actual factorial ablations follow different trajectories and may change the remaining derivatives. Consequently, the empirical factorial interaction need not equal the theoretical joint-path contribution (W1).

Response: Thank you for raising this important question about the relationship between theoretical path decomposition and realized factorial ablation in nonlinear learning systems. We have clarified their mathematical relationship, established conditions for their approximation, and provided a direct empirical comparison in the Digit system. The full analysis is presented in General Response #2.

Theoretical relationship. Path decomposition exactly identifies the joint-path contribution through primitive-arrow deletion while keeping the remaining derivatives fixed along a common trajectory (Proposition 1). In our minimal linear recursion, the parameter masks recover the corresponding learner-response contribution for an initial perturbation u0=exu_0=e_x, because the removed state remains zero. This does not identify the full-state masked matrices with the primitive-arrow-deleted matrices. In nonlinear systems, however, realized factorial ablations may follow different trajectories and evaluate different derivatives, so their interaction need not equal the theoretical joint-path contribution. We have extended Appendix D.4 with sufficient conditions and an error bound showing when realized factorial ablation can approximate the theoretical quantity.

Empirical comparison. We added a direct comparison between the common-trajectory masked-response estimate and realized factorial ablation in the Digit system (Table R5). Without direct coupling, the two quantities agree closely, with equivalence within a 20% margin established in five of six settings. Under stronger coupling, systematic discrepancies appear, with significant differences in five of six settings. These results support the approximation in the tested settings without direct coupling, but not uniform equivalence across configurations.

Implications for experimental interpretation. We therefore distinguish the exact theoretical joint-path contribution from its common-trajectory masked-response estimate and from realized factorial interactions measured across different learning trajectories. The common-trajectory estimate approximates the theoretical quantity subject to the consistency of implemented masks with primitive-arrow deletion, while realized factorial interactions quantify the effects of mechanisms across their actual experimental runs. We interpret their agreement as an empirical finding rather than an exact identity, as clarified in General Response #2 and Appendix N.1.7 (Table 15).

Reviewer_LHRBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper considers a class of recursive self-improving systems which are coupled via multiple feedbacks to a common learner. The authors model this system using the state of the learner, state of an inherited artifact or defect, and state of a responsive environment as a joint dynamical system. They develop a primitive-arrow/path decomposition of finite-horizon Jacobian propagation for this system and argue that factorial ablation (FA) of paths can reveal influence of both feedback mechanisms while leave-one-out ablation (LOOA) does not. They then consider a simple three scalar linear recursion with reinforcement and non-negative routes and derive a correction-budget stability boundary by considering mediating effects through the state of the defect and the state of the environment, including the direct effect between them. Finally, they discuss when internal progress may differ from external performance measured under protection. All the results are evaluated by experiments in linear simulation, a digit self-training loop with a pseudo-label buffer and a curated pool, and a continual test time adaptation experiment on CIFAR-100-C.

Critical strengths

The problem considered here is important for self-training, self-rewarding and agentic systems because conclusions derived from individual feedback loops do not necessarily generalize when such loops are implemented in tandem. The conceptual usefulness of distinguishing between primitives within-round arrows and blocks in a composed Jacobian is noted; specifically, the paper correctly points out that the deletion of state components or inspection of sub-blocks of a composed map need not eliminate all paths through those components.

Under the common trajectory and fixed derivative interpretations of the ablation, the use of factorial (as opposed to leave-one-out) ablations can be motivated from a path decomposition perspective. Under the restricted stationary, reinforcing/nonnegative linear setting considered in this work, the correction budget condition can be interpreted. Furthermore, the distinction between mediated effects and direct defect environment coupling is also useful.

The authors are commended for noting limitations, e.g., the scope restriction of signed and oscillatory systems, nonstationarity, finite window effects, locality vs. globality of conclusions, imperfect linear prediction in the learned setting, and the fact that the CIFAR experiment does not establish RSD.

The digit experiment is strengthened by the use of an identification protocol that estimates quantities from restricted arms prior to the evaluation of some full-loop behavior on held-out seeds.

Critical weaknesses

The central practical claim is not convincingly validated in the main digit experiment. Although the full system exhibits measured local amplification, protected performance continues to improve. This supports the existence of local coupling or perturbation growth but does not establish recursive self-deterioration.

In the stronger-coupling digit experiment, protected performance declines, but measured amplification remains below one even when the linear model predicts instability. Moreover, the claimed post-burn-in regime is not convincingly established as a fixed point of the full dynamics. This raises concerns about the applicability of the proposed stability boundary in learned systems.

The CIFAR experiment does not explicitly identify both feedback mechanisms or test the proposed spectral and factorial predictions. Furthermore, a matched open-loop replay baseline shows similar deterioration, making it difficult to attribute the observed behavior to coupled feedback rather than other factors.

The exact path/factorial identity assumes arrow deletion along a fixed common trajectory. However, in nonlinear learning systems, actual ablations change the learning trajectories, data distributions, and primitive derivatives. Therefore, empirical factorial ablations may not isolate the proposed joint feedback paths as cleanly as the theoretical analysis suggests.

The proposed stability condition relies on stationary local linearization, reinforcing or nonnegative feedback routes, and assumptions connecting unstable learner dynamics to deterioration in protected performance. These restrictions limit its applicability to more general nonlinear, nonstationary, or signed-feedback systems.

Questions for authors

The paper title does not match the submission title in Open Review. Please fix this.

Could you conduct a multi-mechanism experiment with two single-mechanism arms where neither arm has its protected performance deteriorate, yet the system’s protected performance does, and where the boundary between the two states is estimated from the independent single-mechanism arms?

In the case of the strong-coupling digit experiment have you verified that the system is approximately stationary under the unmasked full dynamics? If finite differences were ised (or local Jacobian) to compute the local Jacobians; and can you explain why the system’s protected performance deteriorates but the measured local amplification is less than 1.

Can you clarify the difference between the exact path decomposition along a fixed trajectory, the first order approximation to the effect of an intervention, and the realized factorial ablation after following a different trajectory? Are there any assumptions or diagnostics that justify the interpretation of the empirical factorial ablation as an estimate of the theoretical path contribution?

Can you compare the proposed correction-budget diagnostic with simpler ones like those based on factorial interactions in protected performance, a general local Jacobian fit, or the monitored protected metric, and if the proposed model adds anything to these?

Finally, can you place their proposed stability condition in the context of the classical positive-system, M-matrix, small gain, and dynamical systems analysis and to say whether any of the results presented are new and how far they can be extended to non-stationary, non-reinforcing, and/or non-negative linear systems?

6 comments
AuthorsOpenReview ↗

Response to Reviewer (1/6)

We thank the reviewer for the careful reading and thoughtful comments.

1. Amplification and Deterioration — Concern 1 (W1, W2, Q2, Q3)

Concern: In the original main Digit experiment, the full loop appeared to amplify perturbations while protected performance continued to improve (W1). In the strong-coupling experiment, protected performance declined while measured amplification remained below one, despite the model predicting instability (W2, Q3). The reviewer also asks for a setting in which neither single-mechanism arm deteriorates but the full system does, with the amplification boundary estimated from the single-mechanism arms (Q2).

Response: Thank you for raising these important questions about the relationship between perturbation amplification and protected deterioration in the Digit experiments. The main experimental discrepancies were largely due to the original amplification-rate estimator, which was unreliable near the amplification boundary. The revised measurements resolve the contradiction in the main experiment and establish agreement between measured amplification and protected deterioration under stronger coupling, while the remaining near-boundary behavior requires a finite-horizon interpretation. We summarize our findings below, with the full analysis provided in General Response #1.

Main experiment (W1). The full loop does not amplify perturbations. Its revised amplification rate is 0.989±0.0030.989\pm0.003, closely matching the predicted rate of 0.9900.990 obtained from separately identified single-mechanism responses (Table R3). Both single-mechanism systems attenuate, and the full system also remains on the attenuating side, consistent with protected performance continuing to improve. We have corrected the original claim that this configuration exhibits amplification.

Strong-coupling experiment (W2, Q3). With the revised estimator, the measured amplification rate significantly exceeds one at ξ=0.8\xi=0.8 and 1.01.0 (1.0031.003 and 1.0081.008, respectively), and protected performance declines significantly at both values (Table R4). Neither single-mechanism arm shows significant protected deterioration. At ξ=0.6\xi=0.6, protected performance declines while the measured rate remains just below one (0.9960.996). This does not contradict our theoretical results: the measured rate characterizes local perturbation growth, whereas protected deterioration is assessed over a finite window. Near the boundary, finite-window performance trends need not follow the asymptotic spectral regime, particularly when the learned dynamics satisfy the fixed-point and stationarity assumptions only approximately (General Response #1).

Requested experiment (Q2). The strong-coupling experiment demonstrates the requested deterioration pattern: neither single-mechanism arm shows significant deterioration, while the full system does. However, predicting the strong-coupling boundary additionally requires calibration of direct coupling, whereas the base configuration provides a comparison using separately identified single-mechanism responses. In the tested configurations, the full loop remains attenuating without direct coupling. Thus, the revised experiments establish these two properties in different configurations rather than simultaneously in a single configuration. The base-configuration prediction was also re-identified after the measurement, rather than established prospectively. We clarify these distinctions and the scope of the experimental evidence in General Response #1.

AuthorsOpenReview ↗

Response to Reviewer (2/6)

2. Path Decomposition and Factorial Ablation — Concern 2 (W4, Q4)

Concern: In nonlinear systems, realized factorial ablations may change the trajectories and derivatives of the remaining mechanisms, so the resulting interaction need not equal the joint-path contribution defined along a common trajectory (W4). The reviewer asks us to clarify the distinction between exact path decomposition, first-order approximation, and realized factorial ablation, and the conditions under which factorial ablation can approximate the theoretical joint-path contribution (Q4).

Response: Thank you for highlighting the distinction between exact path decomposition and realized factorial ablation in nonlinear learning systems. A detailed theoretical clarification and empirical comparison are provided in General Response #2.

Theoretical distinction. Path decomposition exactly identifies the joint-path contribution by deleting primitive arrows while keeping the remaining derivatives fixed along a common trajectory (Proposition 1). First-order approximation describes the response to a small perturbation within each arm, but does not eliminate differences between trajectories when mechanisms are removed. In nonlinear systems, realized factorial ablations may therefore follow different trajectories and evaluate different derivatives, so their interaction need not equal the theoretical joint-path contribution. We have extended Appendix D.4 with sufficient conditions and an error bound for this approximation.

Empirical comparison. We have also added a direct comparison between the common-trajectory masked-response estimate and realized factorial ablation in the Digit experiment (Table R5). Without direct coupling, the two quantities agree closely, with equivalence within a 20% margin established in five of six settings. Under stronger coupling, systematic discrepancies arise, with significant differences in five of six settings. We therefore use the common-trajectory masked-response estimate to approximate the theoretical joint-path contribution and interpret realized factorial interactions as empirical quantities rather than exact identities.

3. Composition-Induced Deterioration — Concern 3 (W3)

Concern: The original CIFAR-100-C experiment does not isolate two interacting feedback mechanisms or establish that deterioration is caused by their interaction. Moreover, a matched open-loop replay exhibits similar deterioration, making it difficult to attribute the observed decline to recursive feedback. The reviewer asks for stronger empirical evidence of composition-induced deterioration in realistic AI systems (W3).

Response: Thank you for emphasizing the need to identify interacting feedback mechanisms in realistic AI systems. We address this concern in General Response #3 by clarifying the limitations of the original experiment and introducing a new factorial experiment on RoTTA, a published continual test-time adaptation method.

Original experiment. The original mean-teacher experiment on CIFAR-100-C contains only one feedback route and exhibits similar deterioration under matched open-loop replay, so it cannot establish deterioration caused by composition. We have clarified this limitation in the revised manuscript.

New RoTTA factorial experiment. In the new experiment, we separately control inherited-error feedback through teacher-generated training targets and environment feedback through confidence-based memory admission. The resulting 2×22\times2 factorial analysis identifies a +21.10 percentage-point interaction (95% CI: [15.70, 26.25]; Table R6), showing that the interaction between these feedback routes substantially aggravates deterioration beyond their separate effects.

Interpretation and scope. The inherited-error route already causes deterioration on its own, whereas the environment route has no significant additional deterioration relative to the both-off control. Thus, the experiment demonstrates composition-aggravated deterioration, rather than failure of two individually benign mechanisms. It evaluates the realized factorial interaction, not the spectral amplification boundary, which is investigated separately in the Digit system (General Response #1).

AuthorsOpenReview ↗

Response to Reviewer (3/6)

4. Theoretical Scope and Generalizability — Concern 4 (W5, Q6)

Concern: The proposed stability criterion relies on stationary local linearization and reinforcing or nonnegative feedback, while the connection between amplification and protected deterioration requires additional assumptions. These restrictions may limit the applicability of the theoretical results to more general recursive learning systems (W5). The reviewer also asks how our framework relates to classical positive-system, M-matrix, small-gain, and dynamical-systems theory, which contributions are genuinely new, and how far the results extend beyond the restricted setting (Q6).

Response: Thank you for raising important questions about the scope of our theoretical results and their relationship to classical stability theory. We clarify the assumptions, generalizability, and novelty of our framework in General Response #4.

Theoretical scope and limitations (W5). Our results operate at different levels of generality. Proposition 1 provides an exact finite-horizon path decomposition for differentiable nonlinear and time-varying recursions without requiring a fixed point. The explicit correction-budget amplification criterion (Theorem 3, Corollary 4) applies to the structured linear model under its stated feedback assumptions, while the connection to protected deterioration (Propositions 2, 5) requires additional protected-readout and, for Proposition 5, dominant-mode conditions. These distinctions are summarized in Table R8.

Classical foundations, novelty, and extensions (Q6). Our reinforcing-feedback criterion builds on classical positive-system and M-matrix theory. The contribution lies in identifying and analyzing the feedback mechanisms of recursive learning systems, including their joint propagation and effects on the amplification boundary. Beyond the reinforcing linear setting, our analysis characterizes additional stability conditions for signed feedback, including negative and complex eigenvalue crossings and stabilization through opposing feedback. For nonlinear and nonstationary systems, the exact path-decomposition result remains valid along time-varying trajectories, while the explicit correction-budget boundary applies to the structured local model. A detailed comparison with classical stability theory and other related work is provided in General Response #4 (Table R7).

Comments on the paper

12 comments
AuthorsOpenReview ↗

Summary Comment (1/2)

Dear Reviewers and Area Chair,

We would like to thank all reviewers for your careful evaluations and constructive feedbacks. The concerns centered around four main issues: (1) the relationship between amplification and protected deterioration, (2) the connection between path decomposition and factorial ablation in nonlinear systems, (3) empirical evidence of coupled-feedback deterioration in realistic AI systems, and (4) the novelty and theoretical scope of our framework. We have addressed these concerns through theoretical clarifications, revised measurements, additional experiments, and a more explicit account of our contributions and their applicability. The table below summarizes the reviewers' concerns and how they have been addressed.

Note: Concerns and Actions are numbered separately for each reviewer. W and Q refer to the corresponding Weaknesses and Questions in the original reviews.

ReviewerQuestions / ConcernsHow Concerns Were Addressed
LHRB (R1)Concern1 (W1, W2, Q2, Q3). Relationship between amplification and protected deterioration; inconsistencies in the Digit experiments. Concern2 (W4, Q4). Path decomposition versus realized factorial ablation in nonlinear systems. Concern3 (W3). Evidence of coupled-feedback deterioration in realistic AI systems. Concern4 (W5, Q6). Theoretical scope, novelty, and relationship to classical stability theory. Concern5 (W2, Q3). Stationarity and fixed-point assumptions in the Digit system. Concern6 (Q5). Added value of the correction-budget framework.Action1. Clarified the theoretical conditions linking amplification and deterioration and revised the Digit experiments using a new estimator (General Response #1). Action2. Established approximation conditions and added a direct empirical comparison (General Response #2). Action3. Added a factorial experiment on RoTTA (General Response #3). Action4. Clarified our mechanism-level contributions relative to classical stability theory and distinguished the assumptions and scope of our theoretical results (General Response #4). Action5. Examined the stationarity of the reduced dynamics and the fixed-point assumption. Action6. Clarified the framework's predictive value, supported by results across 77 combined configurations (original estimator; closest agreement away from the boundary).
nqPw (R2)Concern1 (W2, Q1). Whether predicted amplification, measured amplification, and protected deterioration agree in learned systems. Concern2 (W1). Whether realized factorial ablations accurately represent the theoretical joint-path contribution in nonlinear systems.Action1. Revised the amplification-rate estimator and provided experiments showing agreement between predicted and measured rates, as well as settings where amplification and protected deterioration occur together (General Response #1). Action2. Clarified the distinction between exact path decomposition and realized factorial ablation, established approximation conditions, and provided new empirical comparisons (General Response #2).
EzPp (R3)Concern1 (W1). Unclear storyline and presentation. Concern2 (W2). Novelty and scientific significance relative to control theory and existing research. Concern3 (W3, Q1). Limited experimental scale and lack of realistic examples of deterioration caused by interacting feedback mechanisms.Action1. Revised the Introduction to clarify the scientific motivation and narrative. Action 2. Clarified the novelty of our mechanism-level framework relative to classical stability theory and representative lines of related work (General Response #4). Action3. Added a factorial analysis of RoTTA, a published AI system, demonstrating substantial interaction effects between inherited-error feedback and adaptive data selection (General Response #3).
AuthorsOpenReview ↗

Summary Comment (2/2)

In response, the major revisions in the manuscript are as follows:

(1) Amplification and Protected Deterioration — Theoretical Clarification and Revised Experiments.

  • Section 3.4: Consolidated and extended the theoretical results connecting amplification to protected deterioration, including a new proposition for the general recursive framework.
  • Appendix N.1.5–N.1.7: Revised the growth-rate estimator and re-evaluated the Digit experiments. In the base configuration, predicted and measured amplification rates now agree closely (Table 12). Under strong coupling, protected performance deteriorates significantly wherever the measured rate significantly exceeds one (Table 14).
  • Appendix N.1.7: Added estimator validation and diagnostics of the local dynamics.

(2) Path Decomposition and Factorial Ablation — Theoretical Conditions and Empirical Validation.

  • Appendix D.4: Added sufficient conditions and an error bound characterizing when realized factorial ablation approximates the theoretical joint-path contribution.
  • Appendix N.1.6–N.1.7: Clarified the distinction between common-trajectory path decomposition and realized factorial interaction. Added a direct comparison of common-trajectory masked responses and realized factorial responses in the Digit system (Table 15), showing close agreement without direct coupling and systematic discrepancies under stronger coupling.

(3) Coupled-Feedback Deterioration in Realistic AI Systems.

  • Section 5.2: Clarified the limitations of the original CIFAR-100-C experiment and introduced a new factorial experiment on the published RoTTA system.
  • Appendix N.3: Added a controlled comparison of two feedback routes, showing a +21.10 percentage-point factorial interaction (95% CI: [15.70, 26.25]; Table 20 and Figure 11). The results demonstrate that interacting feedback mechanisms can substantially aggravate deterioration in a realistic AI system.

(4) Novelty and Theoretical Scope.

  • Section 2 (Related Work): Clarified our mechanism-level contributions relative to classical stability theory, including the role of positive-system and M-matrix theory in the reinforcing-feedback stability criterion.
  • Appendix B (Table 1): Compared our framework with representative lines of related work in terms of recursive feedback, joint-feedback attribution, deterioration analysis, and coupled-system prediction.
  • Sections 3–4 and Appendices D–J and L: Clarified the different levels of theoretical generality, distinguishing the general path-decomposition framework from the additional assumptions required for the explicit amplification boundary and its relationship to protected deterioration.

We have also revised the Introduction to clarify the scientific motivation and storyline, and further explained the mechanistic and design-level value of the correction-budget framework.

These revisions strengthen the theoretical interpretation, experimental validation, and presentation of our work. All revisions in the manuscript are highlighted in blue. We believe they substantially address the reviewers' main concerns and clarify both the contributions and the scope of our results.

Best regards,

Authors

AuthorsOpenReview ↗

General Response 1 (1/3)

We thank all reviewers for their thoughtful and constructive feedback. We are encouraged that they recognized the importance of studying interacting feedback mechanisms in recursive learning systems (Reviewer LHRB), described our theoretical analysis as “a precise and interpretable account” of why individually attenuating mechanisms can amplify jointly (Reviewer nqPw), and acknowledged the usefulness of the path decomposition and correction-budget framework (Reviewers LHRB, nqPw). We also appreciate their positive assessments of the theoretical soundness (Reviewers nqPw, EzPp), the ability to predict combined-system behavior from separately identified mechanisms (Reviewer nqPw), and the experimental support for our theoretical findings (Reviewer EzPp).

1. Amplification and Deterioration

Response: We thank Reviewers LHRB and nqPw for raising an important question about the relationship between perturbation amplification and protected deterioration. We address this question through theoretical clarification and revised experiments in the Digit system.

1.1 Theoretical Relationship Between Amplification and Deterioration

Amplification describes the growth of perturbations in a recursive system, while deterioration refers to a decline in protected performance. Under appropriate conditions, amplification and protected deterioration characterize the same dynamical regime.

In the general recursive framework, the newly introduced Proposition 2 (extending Appendix J.5–J.8 of the original submission) establishes a local, finite-horizon connection between the protected amplification factor and protected performance change under its protected-readout assumptions, when the reference protected performance is unchanged over the window. This criterion corresponds to ordinary perturbation amplification when the normalized protected curvature is preserved along the propagated direction.

For the minimal solvable model, Proposition 5 (Proposition 4 in the original submission) further establishes that, under its quadratic-readout and dominant-mode conditions, attenuation corresponds to eventual improvement in protected performance, while amplification corresponds to eventual deterioration. Thus, under these conditions, the amplification and deterioration boundaries coincide.

These results establish the theoretical conditions linking amplification to protected deterioration. A decline in protected performance is further identified as recursive self-deterioration (RSD) when it is caused by the feedback mechanisms, as determined by comparison with the learner-only recursion.

1.2 Experimental Validation in the Digit System

The discrepancies in the original Digit experiments came mainly from the amplification-rate estimator. We have revised the estimator and re-measured both experimental configurations. Predicted and measured rates now agree, and wherever the measured rate significantly exceeds one (ξ=0.8\xi=0.8 and 11), protected performance declines significantly. Close to one, the picture is more complex, because there the rate alone does not fix the sign of a finite-window trend. Proposition 2 gives a local finite-horizon relationship under protected-readout and unchanged-reference-performance conditions, whereas Proposition 5 gives an asymptotic result under quadratic-readout and dominant-mode conditions. The fitted reduced one-round spectral radius alone does not establish these hypotheses or determine every finite-window performance trend; we discuss this case explicitly below.

  • Revised growth-rate estimator. The full state of the learned loop has about 4,000 coordinates (learner weights, buffer labels and pool items). We estimate a reduced one-round response map from twin runs: two copies share every random draw, one is perturbed, and we measure their response difference. We report the spectral radius of this fitted reduced map; it is not automatically the growth rate of the full nonlinear dynamics.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.