PaperlayerView exact revision on arXiv

Source-linked AI summary

What You Can't See Is Still What You Learn: A Preregistered Sixty-Society Confirmation That Evidence Masking Drives Compositional Generalization

Narcis Marincat

arXiv:2609.17637v2cs.AIcs.LG

TL;DR

The paper asks whether restricting modules’ evidence access improves generalization to unseen combinations, since full-input access can allow solutions to bypass the intended division of work. It tests masking in a preregistered, five-condition study of four-cell systems that communicate through learned packets. Masking passed the behavioral criteria, while the roles of usable markers and foreign evidence remain unresolved.

  • Problem

    The study asks whether restricting evidence access improves generalization when familiar operations must be combined in previously unseen programs.

    Source: §1 · p. 1

  • Method

    A preregistered study compares evidence masking, ownership markers, and neutral filler in four-cell systems that pass learned continuous packets between modules.

    Source: §1 · pp. 1–2

  • Results

    With markers available, all twelve matched pairs cleared the required margins at both depths, with median paired advantages of +0.846 and +0.859; the unmarked replication also passed.

    Source: §1 · p. 2, §5 · p. 6

  • Takeaways & Limitations

    The tested masking regime has a large advantage, but usable role information and the explanation for the filler condition remain unresolved.

    Source: §1 · p. 2, §7 · p. 10

    The evidence comes from one task world per study, one 0.5B backbone, and within-grammar held-out phrasings; marker-role claims require a shuffled-slot design, and audits do not establish mediation.

    Source: §5 · p. 9, §8 · p. 11

Abstract

from arXiv · show

Restricting what a module can read may improve what a system learns to compute. We test this in a preregistered confirmation with sixty four-cell systems sharing a frozen language-model backbone and communicating through learned continuous packets. Five conditions vary evidence masking, ownership markers, and replacement of foreign evidence with neutral filler (task-irrelevant text of the same token length), across six initialization clusters, each with two data orders, on one fresh task world. With markers available in both regimes, masking improved accuracy on held-out two- and three-operation compositions by median paired differences of 0.846 and 0.859; all twelve pairs cleared the required margins, and the full preregistered behavioral criterion passed. The unmarked replication also passed. No marked global-visibility (G+) system passed the marker-following check, so the effect of usable role information remains unresolved. The filler condition yielded seven full generalizers, but its decomposition criteria were inconclusive. Packet interventions in all eighteen audited masked systems followed the predicted intermediate-value changes on eligible cases; these finite, success-conditioned audits do not establish mediation. The results confirm a large advantage of the tested masking regime, while leaving its finer attribution and generality open. Protocols, results, and checkpoints are public.

1 Introduction

The introduction asks whether restricting modules’ evidence access improves generalization to unseen combinations of familiar operations. In a preregistered confirmation, masking produced a large advantage, while the roles of markers and foreign evidence remain unresolved.

1 Introduction

  • The study tests whether restricting evidence access favors reusable intermediate results on unseen operation combinations, with conclusions bounded to the tested architecture, task world, and training budget.The motivation is that full-input access can permit solutions that bypass the intended division of work.Source: §1 · pp. 1–2
  • Four sequential calls to one shared language model process arithmetic operations or forwarding instructions, passing learned continuous vectors between jointly trained modules.Under masking, each cell reads only its assigned evidence span, the shared question, and its incoming packet; global visibility also exposes other spans.Source: §1 · pp. 1–2
  • With markers available, all twelve matched pairs cleared the required margins at both depths; median paired advantages were +0.846 and +0.859, and the unmarked replication passed.The confirmation used six initialization clusters with two data orders per condition.Source: §1 · p. 2
  • No marked global-visibility system passed the marker-following check, leaving usable role information unresolved; neutral filler yielded seven full generalizers, but its decomposition criteria were inconclusive.The filler replaced foreign evidence with task-irrelevant text of the same token length.Source: §1 · p. 2
  • All eighteen audited masked trajectories passed same-value and counterfactual packet tests on eligible cases, but the audits do not establish that a representation mediates generalization.The tests check whether substituting packets associated with same or different intermediate values preserves or changes the answer as predicted.Source: §1 · p. 2

2 Related work

Related work distinguishes generalization by an existing system from the compositionality of its messages and the ease with which new receivers learn them. The experimental design tests evidence masking in a fresh arithmetic task world using matched conditions, frozen decision rules, and packet-based communication.

2 Related work

  • Compositional generalization concerns familiar components in unseen combinations; SCAN models could struggle when commands combined familiar elements in new ways (Lake & Baroni, 2018).Emergent-communication studies found that successful cooperation need not yield strongly compositional messages, though more compositional codes were easier for new receivers to learn (Kottur et al., 2017; Chaabouni et al., 2020).Source: §2 · p. 2
  • Interchange interventions test what an internal message does by replacing a state and comparing the output with a symbolic prediction; these packet tests cover finite eligible cases, not all inputs or mediation of masking’s advantage.The paper replaces packets and tests whether outputs follow predicted changes in intermediate arithmetic values.Source: §2 · pp. 2–3

3 Experimental design

  • The study builds on the parent masking study in a newly generated arithmetic world, using four-cell systems with shared model parameters and continuous-vector messages.Ownership labels and neutral text replacing other cells’ evidence assess competing explanations.Source: §2 · p. 3
  • The task composes twelve non-identity affine bijections modulo 17; held-out banks contain 18 depth-two and 231 depth-three programs evaluated across all 17 values and four sealed phrasing rotations.These fixed-bank measurements are not independent experimental replicates, and held-out phrasings remain within the generated grammar.Source: §3 · pp. 3–4
  • The learned packet protocol is trained from final-answer supervision rather than supervised intermediate packet targets; depth-two tasks use a forwarding instruction in the final cell.Identity and single-operation episodes also use forwarders in unoccupied operation slots.Source: §3 · p. 3
  • Each society uses a shared frozen-backbone architecture with learned readers, writers, and packet updates; a residual anchor carries the incoming packet into the update.At the output, the fourth-cell packet is mapped into the frozen head’s input space, and residualized logits are used for the fixed answer labels.Source: §3 · p. 3
  • Five conditions vary masked or global evidence access, ownership markers, and neutral filler; the filler arm preserves attention reach and span lengths while replacing foreign evidence tokens.The filler also changes lexical content and may make the genuine evidence span recognizable, so it informs but does not isolate the contribution of foreign content.Source: §3 · p. 4
  • Matched conditions share initialization, training examples, token layout, parameterization, and update budget; preregistered contrasts use frozen conjunction gates and a separate marker-following check.The marker check requires 11 of 12 G+ societies to meet both intact-marker accuracy and counterfactual tracking thresholds before a role-equalized reading is permitted.Source: §3 · pp. 4–5

4 Preregistration and execution

The protocol and numerical decision rules were deposited before seed generation, but execution departed from the planned cohort-wide evaluation embargo. The disclosure also reports missing admission and duplicate-audit records that limit retrospective verification.

4 Preregistration and execution

  • The protocol, code inventory, contrast definitions, and decision rules were deposited before seed generation, with third-party Software Heritage snapshots; the populated manifest was deposited before training.AI reviews were procedural assistance, not journal peer review or independent replication.Source: §4 · p. 5
  • Final checkpoints were evaluated as they completed, so outcomes were inspected before all sixty systems finished and before the correction and verdict; missing records further limit retrospective verification.The frozen evaluator and result records were unchanged by a post-evaluation correction to the derived single-operation denominator.Source: §4 · p. 5

5 Preregistered results

The preregistered results show a large masking advantage on held-out compositions, with the unmarked replication passing as well. The marker-role and neutral-filler controls leave finer attribution unresolved.

5 Preregistered results

  • Median paired advantages were +0.846 at depth two and +0.859 at depth three for R+−G+, with all 12 pairs clearing both margins.The full preregistered behavioral criterion passed.Source: Figure 2 · p. 6, §5 · pp. 6–7
  • The unmarked R−−G− replication passed in all 12 trajectories, with median paired advantages of +0.884 at depth two and +0.917 at depth three.The marked and unmarked comparisons reuse the same initialization–order structure and are not independent replicates.Source: §5 · pp. 6–7
  • R+ and R− met the preregistered practical-redundancy criterion under masking, with median differences of 0.000 at both depths.This was not a formal statistical equivalence test or a causal assay of internal marker use.Source: §5 · p. 7
  • No marked global-visibility trajectory passed the marker-following check, so semantic role rescue remains unresolved.The frozen check required intact-marker accuracy and counterfactual tracking of at least 0.90; observed tracking ranged from 0.038 to 0.060.Source: §5 · pp. 7–8
  • Trajectory-level decompositions hold, but medians do not obey them, so the secondary statistics infer no component share from medians.The reported bootstrap intervals and p-values supplement rather than determine the preregistered decision criteria.Source: Table 4 · p. 8
  • The neutral-filler arm included seven full generalizers, but its preregistered decomposition contrasts were inconclusive.The arm also had four non-acquirers and one shallow-partial trajectory; filler replacement does not isolate foreign-content effects from other consequences.Source: §5 · pp. 8–9
  • All masked trajectories had full-bank all-cut accuracy exactly at chance, 0.0588; this channel-necessity result does not show that packets exclusively carry answer-relevant information.Other competent trajectories were near chance and within the frozen all-cut bound.Source: §5 · p. 9

6 Mechanism and dialect audits (non-gating)

The non-gating audits found value-level intervention fidelity across tested masked trajectories, but not strong canonicality across marked global trajectories. Interoperability between regimes usually failed its bidirectional criterion, and audits were limited to eligible cases and selected trajectories.

Audit design

  • Audits used packet interventions on initially correct recipient cases, with eligible denominators varying across pairs and interfaces; intact replay was not a necessity statistic.The mechanism bank comprised 24 programs, 17 values, and two diagnostic rotations; success-conditioning means its intact rate was 1.000 by construction.Source: §6 · p. 9

Value-level intervention audits

  • All eighteen audited masked trajectories achieved 1.000 same-value preservation and counterfactual following on eligible cases, with deletion at or near chance.The audits cover twelve R+ matrix recipients and six preselected R− self-audits; they do not establish mediation of the masking advantage.Source: §6 · p. 10

Marked-global audits

  • No marked global trajectory met strong canonicality across all three interfaces, and eleven of twelve had an interface meeting the operational context-entanglement indicator.These recipient–interface indicators are not exhaustive society-level semantic diagnoses.Source: §6 · p. 10

N+ audits

  • Among six preselected N+ self-audits, two competent trajectories met strong canonicality at every interface; the evidence supports an association between success and high fidelity, not mediation.Only six of twelve N+ trajectories were audited, and unsuccessful trajectories had small eligible sets.Source: §6 · p. 10

Cross-regime interoperability

  • Eleven of twelve marked pairs failed the bidirectional interoperability conjunction, although individual transfer directions could be substantially stronger than their reverses.The single qualifying pair met the conjunction only at its final interface; poor interoperability does not establish semantically different code types.Source: §6 · p. 10

7 Discussion

The fresh-world confirmation passed both the marker-present contrast and the unmarked replication, while the discussion leaves the mechanisms and broader applicability unresolved. The authors frame evidence boundaries as a design choice to test, not a general recipe.

What the confirmation establishes

  • The confirmation passed the marker-present primary contrast and unmarked replication in full on a fresh world, with every paired comparison exceeding the frozen margin at both depths.The parent cohort and confirmation are reported separately, not pooled.Source: §7 · p. 10

Evidence access, text exposure, and content

  • Seven N+ trajectories met the full-generalizer definition whereas no G+ trajectory did, but filler replacement changed lexical content and the decomposition gates were inconclusive.The discussion treats access to task-relevant foreign evidence as a consistent hypothesis, not an established attribution.Source: §7 · p. 11

Why the markers may have been ignorable

  • Fixed placement of a cell’s own span made ownership markers positionally redundant during training, but the audits did not identify the internal learning mechanism.A shuffled-slot follow-up is intended to remove fixed-slot ownership cues; role following would still need validation before concluding role rescue.Source: §7 · p. 11

The N+ split as an optimization observation

  • One hypothesis is that masking changes which solutions are readily discovered while filler replacement offers a less reliable route; basin structure, shortcut adoption, and span-selective attention were not measured.Seven N+ trajectories reached full generalization, five did not.Source: §7 · p. 11

Implication for modular learning

  • The result motivates testing evidence boundaries as a training design choice, but whether they help larger models or less structured tasks remains empirical.The reported result concerns a small affine-composition world and does not establish a general recipe for language-model systems.Source: §7 · p. 11

Idle-relay pressure

  • Idle-relay pressure is a possible explanation for preserving intermediate values, but the audits did not test its contribution and training did not vary forwarding-cell proportions.The mechanism bank’s depth-three programs used all four cells for a value and three operators.Source: §7 · p. 11

8 Limitations

The reported confirmation is bounded by a small, scaffolded task setting and unresolved methodological and verification limitations. Its design also cannot resolve some observed heterogeneity.

Scope and task setting

  • The task uses a scaffolded relay topology, normalized packet anchor, and singleton/forwarding curriculum, so it does not test composition emerging without that scaffolding.The study also uses one backbone size, within-grammar held-out phrasings, and a 17-state carrier space.Source: §8 · p. 11
  • Evidence comes from one task world per study, a 0.5B-parameter backbone, and within-grammar held-out phrasings rather than free language.The marker manipulation uses two constant strings and is positionally redundant during training; shuffled slots are needed before claims about role information proper.Source: §8 · p. 11

Control and mechanism limitations

  • The N+ filler is lexically distinguishable from real evidence, length matching preserves task-correlated span cues, and decomposition contrasts were inconclusive.The audits are success-conditioned, compare regimes with very different competence, and do not establish canonicality as a mediator.Source: §8 · p. 11

Protocol and verification limitations

  • Primary outcomes became available before cohort training finished, and missing admission transcripts and unretained duplicate-audit outputs limit retrospective verification.At six clusters, the design also cannot resolve the reported C2 heterogeneity or the single partial-interoperability interface.Source: §8 · pp. 11–12

9 Documented follow-up proposals

The documented follow-ups target marker interpretation, idle-relay effects, neutral-filler variability, and whether canonical codes change across generations.

Marker and role information

  • A shuffled-slot marker follow-up would randomize span-to-slot assignment per episode to remove fixed-slot ownership cues and make marker following informative.Functional role following would still need validation before drawing a role-rescue conclusion.Source: §9 · p. 12

Idle-relay account

  • An idle-fraction manipulation would vary the number of evidence-bearing cells to test the idle-relay account directly.The proposal is presented as a follow-up, not as a tested explanation of the reported codes.Source: §9 · p. 12

Filler robustness and iterated learning

  • The proposed N+ fragility study asks what determines which neutral-filler societies reach full generalization, while iterated learning would test whether canonical codes sharpen or drift across generations.The passage distinguishes the dated build-ledger proposals from follow-ups presented without a claim of pre-outcome registration.Source: §9 · p. 12

Reproducibility and artifact release

The release provides the evidence records, code, results, and final checkpoints, while distinguishing these deposits from an end-to-end training reproduction. It also discloses limits on independent verification and records that evaluations used final step-20,000 checkpoints without checkpoint selection.

Reproducibility and artifact release

  • The evidence repository contains the preregistration, code and hashes, manifests, evaluation counts, audit records, verdict archive, and incident ledger.All sixty final step-20,000 checkpoints and a SHA-256 inventory are also deposited; the manuscript source archive rebuilds the paper but does not include checkpoints or claim end-to-end training reproduction.Source: Reproducibility and artifact release · p. 12
  • Historical v5 review records concern earlier drafts and AI-assisted preparation, not independent peer review of this paper.The release also includes a duplicate-execution disclosure; later self-audit copies were not retained, so agreement between competing copies could not be assessed.Source: Reproducibility and artifact release · p. 12
  • Behavioral evaluations and mechanism audits used final step-20,000 checkpoints with no checkpoint selection.Deterministic settings were used, but the passage states they do not substitute for the unperformed comparison between competing audit copies.Source: Reproducibility and artifact release · p. 12

Funding and competing interests

The author reports personal funding, no external research funding, and no relevant competing interests. The author led the research and workflow, while generative AI supported implementation, checks, and manuscript work; these uses were not independent peer review or replication.

Funding and competing interests

Use of generative AI

  • The author originated the research ideas and design and developed and oversaw the experimental workflow, including monitoring outputs.Source: Use of generative AI · p. 12
  • Generative AI supported coding, pipeline operation, result checks, and manuscript drafting and revision, but did not provide independent peer review or replication.The author takes responsibility for the design, implementation, analysis, manuscript, and claims.Source: Use of generative AI · pp. 12–13

A Protocol and integrity records

The protocol froze code, thresholds, definitions, and evaluation components before seed generation and training. Machine admission, a one-shot verdict procedure, and disclosed verification boundaries document how the cohort was run and assessed.

Protocol and integrity records

  • Code, gate thresholds, contrast definitions, classifiers, grammars, filler banks, and the verdict script were hash-inventoried before seed generation.The freeze deposit was archived before the master seed was drawn.Source: A Protocol and integrity records · p. 14
  • World seeds were derived from a committed master string, with candidate worlds accepted only by frozen structural checks and previously exposed operator seeds excluded.The populated pre-run manifest was deposited before training was permitted.Source: A Protocol and integrity records · p. 14
  • An AI protocol referee supplied written rulings incorporated as binding constraints, and training required an explicit GO ruling.This was AI-assisted methodological review, not external institutional replication or conventional journal peer review; the author retained responsibility for the protocol and claims.Source: A Protocol and integrity records · p. 14
  • Every rented GPU instance had to pass hardware, artifact-integrity, golden-retrace, and mask-verification gates before cohort training.The golden retrace tests exact agreement on a short reference prefix, not equivalence for every computation.Source: A Protocol and integrity records · p. 14
  • Mask checks tested leakage, source-span liveness, and packet-reader liveness, but retained test transcripts were unavailable for independent re-verification.Per-society three-update fingerprints were recovered for all sixty societies, while one cross-host bit-exact comparison covered only the first three updates and was not an end-to-end reproduction.Source: A Protocol and integrity records · p. 14
  • The verdict script ran once under an exclusive lock on the compiled results and corrected manifest, with its process record archived.The post-evaluation, pre-verdict denominator correction and deviation from the cohort-wide evaluation embargo are documented in Appendix B.Source: A Protocol and integrity records · p. 14

B The manifest correction, in full

A stale two-rotation denominator in the pre-run manifest triggered a fail-closed compiler stop and was corrected to match the frozen four-rotation evaluation specification. The record also documents the correction’s timing and the early execution of primary evaluations relative to the last checkpoint.

B The manifest correction, in full

  • The compiler aborted when the manifest expected 1,224 single-operation diagnostic cases but the frozen evaluator specified 2,448.The manifest expression retained a stale two-rotation factor instead of the specified four rotations.Source: B The manifest correction, in full · p. 15
  • The correction changed only two derived manifest fields from 1,224 to 2,448; the stale value had not been used in training, case selection, or score normalization.All sixty stored S1 scores used the frozen four-rotation bank; the referee's ruling was conditional on submitted facts, not independent certification of later execution.Source: B The manifest correction, in full · p. 15
  • The correction deposit preceded the verdict process start by 33 seconds, and the verdict and bundle were deposited at 20:00:28 BST.The chronology uses recorded Git and process timestamps; the narrative timestamps were approximate.Source: B The manifest correction, in full · p. 15
  • 59 of 60 primary evaluations occurred before the last checkpoint completed, contrary to the preregistered evaluation timing.Evaluations were deterministic, frozen, and final-checkpoint-only; no outcome-dependent stopping, arm selection, audit selection, or scientific protocol amendment is reported.Source: B The manifest correction, in full · p. 15
  • After the correction, the unchanged compiler processed the unchanged society records, and the separate verdict program ran once under the exclusive lock.The first compiler invocation had aborted before computing a verdict, contrast decision, or phenotype classification.Source: B The manifest correction, in full · p. 15

Related papers

Loading 2609.17637v2…