PaperlayerView exact revision on arXiv

Source-linked AI summary

Can Neural Networks Learn by Experimenting on Themselves? Self-Interventional Learning from Functional Consequences to Predictive Self-Knowledge

Michał Tomaszewski

arXiv:2608.14894v1cs.LG

TL;DR

Neural networks rarely learn predictive models of their own functional organization. This paper introduces Self-Interventional Learning, showing that self-generated intervention experience can improve prospective prediction and guide later action, while remaining incomplete and not universally better than direct strategies.

  • Problem

    The paper addresses the limited treatment of a learner’s own functional organization as something it can learn about from intervention evidence.

  • Method

    Self-Interventional Learning lets a neural system perturb its functional structure, observe consequences, learn a predictive self-model, and use it to guide later structural action.

  • Results

    Across experiments, SIL learned predictive self-knowledge that generalized to unexecuted interventions, improved with experience, and reduced decision regret when coupled to action.

  • Takeaways & Limitations

    The evidence supports self-intervention as a framework for acquiring useful predictive knowledge about a network’s functional organization and guiding later action.

  • Takeaways & Limitations

    Structural self-knowledge remained incomplete, with synergy recovered poorly and model-guided action not significantly outperforming direct empirical memory.

Abstract

from arXiv · show

Machine-learning systems usually model external data, while their internal functional organization is analyzed by external observers. This work introduces Self-Interventional Learning (SIL), in which a neural system perturbs its own functional structure, observes consequences, learns a predictive self-model, generalizes to unexecuted interventions, and uses predictions to guide later structural action. In a construction-known synthetic system, SIL recovered critical structure, redundancy, and replaceability, while synergy was not reliably recovered. Across 30 fresh confirmatory seeds, increasing the pairwise intervention budget from 4 to 56 reduced held-out prediction error from 0.0335 to 0.0148 and increased Spearman correlation from 0.629 to 0.883. In a matched ablation, preserving the correct intervention--consequence mapping reduced prospective prediction error by 81.3%, while using the same learned self-model for action reduced normalized regret by 31.7% relative to ignoring it. However, model-guided action did not significantly outperform a direct empirical-memory policy, and powered CIFAR-10/ResNet validation showed no robustness advantage over equal-budget direct repair search. These results support SIL as an intervention-driven framework for learning predictive knowledge about a network's own functional organization, while showing that the self-model remains incomplete and is not universally superior to simpler direct strategies.

1 Introduction

Self-Interventional Learning (SIL) makes a neural system’s own functional structure the target of controlled intervention, using observed consequences to learn predictive self-knowledge and guide structural action. Experiments show generalization to unexecuted interventions and improved prediction with more evidence, while retaining failures for synergy and no universal advantage over direct search.

  • Applied boundary: SIL is presented as a framework for predictive self-knowledge rather than a universally superior robustness method, because powered CIFAR-10/ResNet validation found approximately zero learned-versus-direct performance advantage.Predictive knowledge transferred more reliably than superiority over equal-budget direct search.
  • Confirmatory self-knowledge learning curve: From 4 to 56 pairwise interventions, mean held-out error fell from 0.0335 to 0.0148, while Spearman rank correlation increased from 0.629 to 0.883.Across 30 fresh confirmatory seeds, increasing the nested intervention budget also improved calibration while singleton evidence and the held-out evaluation set remained fixed.
  • Causal intervention–knowledge–action ablation: Correctly preserving the intervention–consequence mapping reduced held-out mean absolute error from 0.0740 to 0.0138 relative to the permuted control.The negative control preserved the quantity and marginal distribution of pairwise evidence but permuted intervention identities and consequences.
  • Framework: SIL uses controlled interventions on a learner’s own functional structure to generate experience for a predictive self-model, which is evaluated on unexecuted interventions and can guide structural action.Interventions may remove, suppress, replace, or jointly manipulate components; their functional consequences provide the training signal.
  • Ground-truth structural test: The construction-known synthetic test strongly recovered redundancy and replaceability, but synergy was not reliably recovered.The protocol separately evaluates criticality, redundancy, replaceability, interaction recovery, structural roles, and genuinely unexecuted interventions.

2 Related Work

SIL combines self-modeling, intervention-based analysis, structural adaptation, and causal reasoning around a distinct setting: a neural learner experimentally models its own functional substrate. Unlike prior approaches that use self-models, interventions, or internal prediction separately, SIL learns predictive self-knowledge from self-generated consequences, tests generalization to unexecuted interventions, and uses that knowledge to guide action.

  • SIL’s distinctive setting: SIL targets the learner’s own functional substrate, observes intervention consequences itself, learns a predictive self-model, tests unexecuted interventions, and guides later structural action.Its relevant unknowns include criticality, redundancy, substitutability, interaction, and prospective functional consequence under structural intervention.
  • Self-modeling: Unlike embodied-robot self-modeling, SIL models internal functional organization rather than external morphology or physical dynamics.Prior robotics work showed self-models could support prediction, planning, adaptation, and recovery from structural change.
  • Interpretability: Unlike interpretable-by-design networks, SIL learns a predictive model of how its own function changes under intervention rather than a human-facing explanation of input–output decisions.Accordingly, the experiments assess predictive accuracy, calibration, structural-relation recovery, and decision regret instead of explanation plausibility or concept alignment.
  • Mechanistic interpretability: Unlike mechanistic interpretability, SIL makes intervention evidence part of the learner’s own process rather than an external analyst’s explanatory procedure.SIL evaluates whether self-models generalize to intervention consequences not observed during self-discovery.
  • Structural adaptation: Unlike pruning and dynamic-sparsity methods, SIL uses controlled structural-intervention outcomes as training data for a model of the network’s own intervention response.Those neighboring methods generally target compression, efficient optimization, or direct task performance.
  • Contribution: SIL’s contribution is the combination of self-interrogation, consequence-based predictive self-knowledge, generalization, and action guidance, while explicitly limiting claims of universal superiority.The paper distinguishes this formulation from prior uses of self-models, interventions, causal reasoning, and internal prediction in isolation.

3 Self-Interventional Learning

Self-Interventional Learning (SIL) uses controlled interventions on a learner’s own functional substrate, observed consequences, and a predictive self-model to generalize beyond executed interventions. The framework tests experience-dependent improvement and knowledge-guided structural action while explicitly limiting claims of self-awareness or universal superiority over direct strategies.

  • Self-Interventional Learning: SIL interventions modify the learner’s functional substrate without changing the external input distribution, including lesioning, joint lesioning, substitution, and other predeclared transformations.Synthetic discrepancies use unlabeled probe inputs, so self-discovery does not require task labels.
  • Predictive self-model: The predictive self-model maps descriptions of self-interventions to functional consequences and is evaluated on held-out interventions whose true consequences were withheld before model freeze.Predictive self-knowledge is assessed using preregistered error, ranking, and calibration metrics.
  • Relational self-knowledge: SIL distinguishes scalar importance, relational self-knowledge from joint or substitutive interventions, and predictive self-modeling that generalizes to unexecuted interventions.Under the synthetic construction, positive joint effects indicate redundancy-like dependence, negative effects indicate complementary or synergistic structure, and replaceability uses directed substitutions.
  • Experience-dependent improvement: Experience-dependent learning requires predictive quality to improve systematically as intervention budget increases, with trends evaluated across independent seeds rather than intervention-level observations.Nested experience sets define the budget comparison, and prediction error is expected to decrease within seeds.
  • Framework boundaries: SIL is defined by jointly testable self-intervention, consequence observation, predictive modeling, held-out generalization, experience-dependent improvement, and knowledge-guided action, not self-awareness or universal strategic superiority.A direct intervention memory may remain optimal in approximately additive regions, while learned relational models target generalization across unexecuted interventions.
  • Empirical program: The empirical program tests structure recovery, budget-dependent predictive improvement, and intervention–knowledge–action coupling, while CIFAR-10/ResNet replication provides an external boundary on repair-selection advantages.The framework also transfers to Kintsugi/CPPV structural repair experiments, but the powered CIFAR-10/ResNet replication found no corresponding performance advantage over equal-budget direct search.

4 Experimental Program and Methods … 4.4 Conditional Prospective Protection Value

The experimental program separated applied structural adaptation from the broader claim that learners can acquire predictive self-knowledge through self-intervention. It used label-free, firewall-controlled interventions, matched structural comparisons, and a conditional repair-value protocol with independent replication.

  • 4.1 Study logic: The program tested applied structural adaptation separately from progressively stronger requirements for Self-Interventional Learning, including recovery of learner-hidden construction-known structure.The Kintsugi/CPPV lineage provided the applied image-classification instantiation, while three mechanistic experiments isolated stronger SIL requirements.
  • 4.1 Study logic: Confirmatory inference treated each seed as an independent replication unit, with fresh seeds used unless a frozen protocol specified a matched within-seed comparison.Pilot results were restricted to technical validation, protocol promotion, and pre-confirmation defect correction.
  • 4.2 Common intervention and firewall principles: After supervised task pretraining, self-intervention, self-model fitting, structural selection, and repair selection used unlabeled buffers or probe inputs without final labels or evaluator-only roles.This firewall separated ordinary task learning from self-interventional learning and self-discovery.
  • 4.2 Common intervention and firewall principles: Held-out structural interventions were genuinely unexecuted before model or policy freezing, enabling evaluation beyond outcomes merely withheld from a fitted regressor.Where structural ground truth existed, acquisition policies used only learner-visible descriptors, and intervention catalogs verified zero discovery–challenge overlap.
  • 4.3 Applied instantiation: Kintsugi structural adaptation: Fashion-MNIST Kintsugi training used an ElasticCNN with reserve capacity, followed by an unlabeled SIL stage performing 80 adaptation interventions after 16 warm-up episodes.SIL-K explored four predeclared intervention families, while final damage families were reserved for evaluation.
  • 4.3 Applied instantiation: Kintsugi structural adaptation: Importance-based regrowth received exactly 80 measured probe interventions and the same seven structural opportunities as SIL-K, alongside frozen, dropout, damage, distillation, and random-regrowth controls.The information budget of the strongest direct structural comparator was explicitly matched.
  • 4.4 Conditional Prospective Protection Value: CPPV estimated prospective candidate-repair value using four reversible repair trials per checkpoint, for a maximum of 28 repair micro-trials, with layer-balanced random probing.Candidate feasibility required a Jensen–Shannon discrepancy threshold of 0.002 on clean outputs; learned selection maximized bG−0.25bσG, while Direct CPPV selected among measured candidates.
  • 4.4 Conditional Prospective Protection Value: The replicated Fashion-MNIST CPPV protocol used 35 new seeds and compared learned CPPV with equal-budget Direct CPPV and SIL-K singleton-only under a four-test Holm family.Primary endpoints were mean damaged accuracy and mean retention, and earlier phase data were not pooled into replication statistics.

4.5 External CIFAR-10/ResNet validation · 4.6 Synthetic functional system shared by Experiments A–C

The paper externally validated the CPPV comparison on CIFAR-10/Elastic-CifarResNet20, while Experiments A–C used an anonymized, construction-known modular system with predefined functional roles. The synthetic protocol measured intervention consequences without exposing labels and audited each seed against predeclared role inequalities.

  • 4.5 External CIFAR-10/ResNet validation: CIFAR-10/Elastic-CifarResNet20 validation retained CPPV logic, held-out evaluation, and matched learned-versus-direct comparison, but an initial transfer run revealed an architecture-specific consolidation failure.Global consolidation and BatchNorm running-statistic changes caused reversible repairs to violate the frozen clean-function safeguard; this run was retained as a transfer diagnostic.
  • 4.5 External CIFAR-10/ResNet validation: The final powered external replication used 82 fresh seeds, equal microtrial budgets and structural opportunities for Learned and Direct CPPV, and the same two primary endpoints under frozen multiplicity.The endpoints were mean damaged accuracy and mean retention.
  • 4.6 Synthetic functional system shared by Experiments A–C: Experiments A–C used a construction-known modular neural system whose organization was hidden from the learner.The passage introduces the shared synthetic system for all three experiments.
  • 4.6 Synthetic functional system shared by Experiments A–C: A fixed modular circuit mapped x to a binary logit, while task labels only verified that the intact system defined a nontrivial classification problem.Those labels were not part of the learner’s available information, as specified in the following passage.
  • 4.6 Synthetic functional system shared by Experiments A–C: SIL never received task labels, and the 24 functional components were anonymized by a seed-specific permutation into public identifiers U00, . . . , U23.This concealed the construction’s component identities from the learner.
  • 4.6 Synthetic functional system shared by Experiments A–C: The construction included critical components, redundant pairs, complementary/synergistic pairs, replaceable target–donor pairs, peripheral components, and neutral components.Critical components made large unique contributions; redundant modules tolerated either surviving member, whereas complementary modules required both members to remain active.
  • 4.6 Synthetic functional system shared by Experiments A–C: Functional consequence for an intervention I was measured on an unlabeled probe set using the intact and intervened output probabilities p0 and pI.The pair-lesion analysis used a separate interaction quantity J.
  • 4.6 Synthetic functional system shared by Experiments A–C: Positive J signaled redundancy and negative J signaled complementarity/synergy, while an oracle-only construction audit verified each seed against predeclared inequalities before learning began.Seeds failing those inequalities were technical protocol invalidations and could not be silently replaced.

4.7 Experiment A: synthetic structural ground truth · 4.8 Experiment B: intervention-budget learning curve

Experiments A and B tested whether self-interventions could recover functional organization and how increasing pairwise experience affected relational prediction. Experiment A used controlled structural discovery, while Experiment B varied pairwise intervention budgets with fixed singleton experience.

  • 4.7 Experiment A: synthetic structural ground truth: Experiment A evaluated recovery of known functional organization through singleton lesions, pair lesions, and directed substitution interventions.All 24 singleton lesions were measured, while pairwise discovery used a structural-blind budget based on learner-visible evidence.
  • 4.7 Experiment A: synthetic structural ground truth: The challenge holdout contained one true redundant pair, one true synergy pair, and one true replaceable relation for every seed.Additional pair lesions and substitution candidates were held out by a seed-stable hash and executed only after freezing the SIL model.
  • 4.7 Experiment A: synthetic structural ground truth: Seven methods compared structural discovery using random, magnitude-, gradient-, singleton-, and SIL-based evidence.The comparisons included random ranking, weight magnitude, activation magnitude, gradient/Taylor importance, direct singleton ablation, a SIL singleton model, and SIL with selected pairwise evidence.
  • 4.7 Experiment A: synthetic structural ground truth: Confirmatory Experiment A used 20 fresh seeds, 56 pair lesions, eight substitutions, and 88 discovery interventions per seed.The confirmatory design also used 12,000 probe inputs and a 60,000-sample construction audit.
  • 4.8 Experiment B: intervention-budget learning curve: Experiment B isolated relational self-interventional experience by fixing 24 singleton interventions and increasing pairwise experience across predeclared budgets.The B = 0 condition served as a descriptive singleton-only anchor, and no substitution interventions were used.
  • 4.8 Experiment B: intervention-budget learning curve: Each Experiment B budget was a strict prefix of one deterministic random ordering, so budget differences reflected added experience rather than different samples.A secondary informed policy used only learner-visible pre-pair information and fixed weights of 0.40/0.25/0.20/0.15, without adapting to observed pair outcomes.
  • 4.8 Experiment B: intervention-budget learning curve: Experiment B measured held-out pair MAE and held-out pair Spearman correlation, testing negative MAE and positive Fisher-z Spearman slopes over log2 B.Confirmatory inference used 30 fresh seeds and one-sided Wilcoxon signed-rank tests with a two-test Holm family.

4.9 Experiment C: intervention–knowledge–action causal ablation · 4.10 Reproducibility and protocol provenance

Experiment C tested whether intervention-derived self-knowledge causally improves prospective structural action by separating intervention, prediction, and action links under a strict commit barrier. Reproducibility relied on protocol-locked repositories, validated execution, and retained protocol locks across Experiments A–C.

  • 4.9 Experiment C: intervention–knowledge–action causal ablation: Experiment C used 80 discovery interventions per seed and evaluated nine two-component reconfiguration episodes with six candidate pairs each.The self-interventional conditions received 24 singleton and 56 random-nested pair interventions per seed.
  • 4.9 Experiment C: intervention–knowledge–action causal ablation: C0–C3 broke different links in the SIL chain, contrasting random choice, direct singleton-additive scoring, prediction-ignored action, and full model-guided action.C1 selected the candidate minimizing C(i) + C(j), whereas C2 used the fitted self-model but ignored its predictions.
  • 4.9 Experiment C: intervention–knowledge–action causal ablation: Permuting the 56 pair consequences preserved sample count, model class, singleton evidence, and marginal targets while destroying pair-identity-to-consequence mapping.This predictor-only negative control isolated the intervention-to-knowledge link.
  • 4.9 Experiment C: intervention–knowledge–action causal ablation: A strict commit barrier froze predictors, stored held-out predictions, committed and hashed all action choices, and verified C0/C2 identities before outcome revelation.Unique held-out candidate interventions were executed only after these checks, with oracle evaluation last.
  • 4.9 Experiment C: intervention–knowledge–action causal ablation: Confirmatory inference used 30 fresh seeds, 12,000 probe inputs, and a 60,000-sample construction audit, testing three preregistered comparative hypotheses.The hypotheses concerned true versus permuted self-model MAE, C3 versus C2 regret, and C3 versus C1 regret, analyzed with one-sided paired Wilcoxon tests in a three-test Holm family.
  • 4.10 Reproducibility and protocol provenance: Experiments A–C ran from protocol-locked repositories containing machine-readable protocols, validators, deterministic resume state, automated summaries, statistical audits, and result packagers.The repositories retained corresponding protocol locks with result archives.
  • 4.10 Reproducibility and protocol provenance: The protocol identifiers were SIL-A-SYNTHETIC-GT-1.0.0, SIL-B-LEARNING-CURVE-1.0.0, and SIL-C-INTERVENTION-KNOWLEDGE-ACTION-1.0.0.These identifiers linked the three experiments to their retained protocol locks and result archives.

5 Statistical Discipline and Evidence Freezing

The analysis froze inferential units, comparison families, information barriers, and decision protocols to prevent pseudoreplication, leakage, multiplicity inflation, and post-hoc rescue. It retained negative findings as part of the study’s claim boundary.

  • Inferential units: Seed-level metrics, rather than individual interventions or episodes, formed the independent inferential units, with paired seed-level differences used when experimental components were shared.This avoided pseudoreplication from treating multiple interventions generated by one network instance as independent.
  • Frozen comparisons: Primary comparisons used frozen nonparametric tests and predeclared directional hypotheses only when ordered predictions were specified before confirmatory outcomes.Otherwise, comparisons remained two-sided within the corresponding lineage.
  • Multiplicity control: Holm correction controlled family-wise α = 0.05 within predeclared scientific families, while CIFAR-10/ResNet retained a separate frozen primary family.Unrelated tests, seeds, and p-values were not pooled across experimental lineages.
  • Multiplicity control: Secondary endpoints and families could refine mechanisms or scope but could not convert a failed primary gate into a positive confirmatory conclusion.Severe-damage, calibration, acquisition, and structural-role analyses were explicitly barred from rescuing failed primary claims.
  • Information firewalls: Three firewall levels withheld task labels, held-out damage families or evaluator-side roles, and unexecuted held-out consequences from adaptation and self-modeling before model freeze.Experiment C additionally required action decisions to be committed and hashed before consequence acquisition.
  • No-rescue rules: Negative evidence was retained after frozen protocols and a hard stop, including failures involving synergy, informed acquisition, Full SIL versus direct empirical memory, dropout, and the powered CIFAR-10 Learned-versus-Direct comparison.Confirmatory seeds, endpoints, thresholds, and multiplicity families were not altered after outcome inspection.

6 Results

Across the experiments, SIL learned predictive relational structure and improved with more intervention data, especially for redundancy, but synergy recovery and broad performance superiority remained limited. Model-guided action helped relative to ignoring the model, yet did not significantly outperform direct empirical memory or direct repair search.

  • External validation: Powered CIFAR-10/ResNet replication found no robustness advantage over equal-budget direct repair search, despite positive predictive self-knowledge signals.Learned-minus-Direct damaged accuracy was only +0.033 percentage points, while mean Protection Spearman remained approximately 0.688 and was positive in 81/82 seeds.
  • Experiment A: Pairwise SIL recovered critical-unit ranking perfectly and substantially improved redundancy prediction over the direct singleton-additive baseline.Mean AUROC was 1.000 for criticality, while redundancy AP was 0.5411 versus 0.01087 for the direct singleton-additive baseline.
  • Experiment A: Synergy remained the principal structural failure, with mean synergy AP only 0.01645 and the preregistered hypothesis failing after Holm correction.Precision and recall at the true-k operating point were zero, and the mean preregistered difference was only +0.00558.
  • Budget scaling: 56 pairwise interventions reduced held-out MAE from 0.03351 at B = 4 to 0.01478 at B = 56, while held-out Spearman increased from 0.629 to 0.883.Expected calibration error also fell from 0.02614 to 0.00662 across the same budget range.
  • Experiment C: Preserving intervention–consequence semantics reduced held-out MAE from 0.07396 to 0.01383, and using model predictions reduced normalized regret from 0.38256 to 0.26140.These matched contrasts isolated the causal contributions of predictive self-knowledge to prediction and action.
  • Experiment C: Model-guided action did not significantly outperform direct empirical memory, with mean regret 0.26140 for Full SIL versus 0.28588 for the singleton-additive policy.The preregistered C3-minus-C1 effect was −0.02448, with Holm p = 0.2449.

7 Discussion

The discussion supports Self-Interventional Learning as an operational framework in which intervention experience yields predictive knowledge of relational functional structure and can guide later action. The evidence also shows important limits: synergy remains difficult, acquisition heuristics do not solve experiment selection, and model-guided action is not universally better than direct empirical strategies.

  • Learning from intervention experience: Increasing observed pairwise self-interventions systematically improved held-out prediction error, ranking quality, and calibration across fresh confirmatory seeds.Experiment B interprets these changes as evidence for an experience-dependent self-model rather than a fixed diagnostic.
  • Using self-knowledge: 81.3% lower prospective prediction error followed from preserving intervention–consequence mappings, showing that useful self-knowledge depends on relational correspondence rather than marginal evidence alone.The matched control permuted consequences while preserving evidence amount and marginal target distribution.
  • Using self-knowledge: 31.7% lower normalized regret resulted when the learned self-model guided action instead of being ignored, but Full SIL did not significantly outperform direct empirical memory.Relational knowledge was especially effective for redundancy-sensitive decisions, while direct singleton-additive memory remained stronger on matched-general cases.
  • Applied transfer: Predictive protection knowledge and extrapolation transferred to powered CIFAR-10/ResNet, but performance superiority over equal-budget direct CPPV was approximately zero.The discussion separates mechanism transfer from performance superiority; on Fashion-MNIST/ElasticCNN, the predictive mechanism yielded a replicated performance advantage.
  • Structural self-knowledge: Pairwise SIL recovered relational structure beyond scalar importance, especially redundancy and replaceability, while synergistic structure remained unreliable.Pairwise SIL recovered the sign of 58/60 true redundant interactions, whereas synergy-sensitive structure and decisions remained difficult.
  • Open limitations: The prespecified informed-acquisition heuristic did not improve sample efficiency, leaving selection of interventions by expected information value, safety, cost, and decision relevance as an open problem.The primary random-nested acquisition policy produced a strong learning curve, but the heuristic was often worse.

8 Limitations and Future Work

The study’s conclusions are bounded by synthetic ground truth, incomplete synergy modeling, restricted interventions, unresolved acquisition strategy, transfer limits, and computational scaling concerns. Future work should address these boundaries while clarifying that SIL’s self-knowledge is predictive functional knowledge, not consciousness or introspection.

  • Scope of evidence: The strongest causal tests used a constructed 24-component system, so structural recovery may not generalize to modern large-scale networks.Applied Fashion-MNIST and CIFAR-10 experiments provide external evidence, but do not establish equivalent role-taxonomy recovery at scale.
  • Interaction modeling: Synergy remained the most consistent weakness: its preregistered hypothesis failed, interaction signs were poorly recovered, and synergy-sensitive decisions remained difficult.The experiments do not distinguish whether the failure arose from representation, learner capacity, acquisition distribution, consequence definition, or constructed synergy geometry.
  • Decision strategy: Direct empirical strategies remained competitive: singleton-additive prediction had the lowest global held-out MAE, while Full SIL did not significantly outperform direct empirical memory.In Experiment C, Full SIL had lower mean regret but lower exact-best and top-2 action rates.
  • Intervention acquisition: More self-interventional experience improved predictive self-knowledge, but informed acquisition did not improve sample efficiency over random nested exploration.The work therefore demonstrates learning from interventions, not an optimal intervention-selection strategy.
  • Intervention space: The final causal experiment used discrete two-component reconfiguration and mainly lesions or substitutions, leaving continuous, heterogeneous, and multi-scale interventions untested.Real systems also permit parameter edits, routing changes, activation manipulations, and module replacement.
  • Architecture and domain transfer: On CIFAR-10/ResNet, Learned CPPV did not outperform equal-budget Direct CPPV in the powered N = 82 replication.Candidate-conditioned protection remained predictable and model-based selection extrapolated to unmeasured candidates, supporting predictive-mechanism transfer more strongly than downstream-performance transfer.
  • Scale and computational economy: Foundation-scale deployment remains unresolved because self-interventions may be expensive, unsafe, or irreversible, requiring surrogate evaluation, reversible interventions, and uncertainty-aware stopping.The mechanistic experiments were intentionally small to permit fresh-seed replications, strict intervention firewalls, and construction audits.
  • Future work: Future work should improve interaction representations, active self-experiment design, direct-memory versus learned-prediction selection, intervention heterogeneity, multi-scale self-models, and decision tasks valuing relational knowledge.The proposed extensions follow directly from the observed boundaries rather than attempts to improve frozen results.

9 Conclusion

The study supports a qualified affirmative answer that neural networks can learn about their own functional organization through controlled self-intervention. SIL learned prospective intervention–consequence mappings, generalized beyond executed interventions, and improved decision-making, while remaining incomplete and not universally superior to simpler strategies.

  • Evidence for SIL: Controlled interventions on a neural system’s functional substrate produced consequences that SIL learned prospectively, recovering nontrivial relational structure.The resulting self-model generalized to interventions not executed during self-discovery.
  • Evidence for SIL: More self-interventional experience systematically improved prediction, and coupling model predictions to action reduced decision regret.These results establish prospective use of the learned self-model rather than post-hoc explanation alone.
  • Limits: Structural self-knowledge was incomplete: redundancy and replaceability were recovered more reliably than synergy.The conclusion therefore supports partial, rather than complete, understanding of the learner’s functional organization.
  • Limits: Model-guided action outperformed ignoring the same self-model but did not significantly outperform direct empirical memory, limiting claims of universal superiority.A simple informed acquisition heuristic also failed to improve sample efficiency.

Data and Code Availability

The submission includes reproducibility materials for Experiments A–C, while using publicly available Fashion-MNIST and CIFAR-10 data and excluding unnecessary large checkpoints from the lightweight archive.

  • Frozen protocols, runnable implementations, and lightweight confirmatory artifacts for Experiments A–C accompany the reproducibility appendix.
  • Fashion-MNIST and CIFAR-10 are used as publicly available benchmark data sets.
  • Large model checkpoints are intentionally excluded because they are not required for the synthetic A–C analyses.
Loading 2608.14894v1…