PaperlayerView exact revision on arXiv

Source-linked AI summary

The Predictability Bias of Joint-Embedding Predictive Learning in Medical Imaging

Fabian Drexel, Bernhard Kainz, Jakob N. Kather, Daniel Rueckert, Lisa C. Adams, Keno K. Bressem, Cosmin I. Bercea

arXiv:2610.13237v1eess.IVcs.CVcs.LG

TL;DR

In medical imaging, recurrent anatomy can be predictable from context while sparse pathology may remain unresolved, and downstream benchmarks do not reveal which local information the predictive objective preserves. The paper trains TAP-JEPA and replays its predictive pathway to audit that distinction. Pathology shows a larger normalized Skill shortfall than anatomical identity in both evaluated datasets, while the analyses are scoped to a large-scale 3D CT JEPA.

  • Problem

    Transfer benchmarks show what representations support downstream but not which local information the predictive objective preserves, a distinction that matters because CT anatomy is recurrent while pathology can be sparse and unpredictable from context.

    Source: §1 · p. 2

  • Method

    The paper trains TAP-JEPA, then replays its exact training mask sampler, teacher, predictor, and loss to trace annotated pathology through the predictive pathway.

    Source: §1 · p. 2

  • Results

    Pathology has a larger normalized Skill shortfall than anatomical identity in ATLAS (36.7% versus 11.2%) and ReX (31.8% versus 6.2%).

    Source: §5.3 · p. 6

  • Takeaways & Limitations

    The results support the conclusion that contextual predictability is not a sufficient proxy for clinical relevance in medical representation learning.

    Source: §6 · p. 10

    The analyses characterize predictability bias in a large-scale 3D CT JEPA, with downstream interventions measuring direct dependence on final-layer pathology tokens and context-side analysis using full replacement on shared surrounding targets.

    Source: §5.6 · p. 9

Abstract

from arXiv · show

Joint-embedding predictive architectures (JEPAs) learn by predicting teacher representations of masked regions from visible context. Point prediction summarizes the conditional target distribution and cannot reproduce variation that the context does not resolve. We call this constraint predictability bias. In computed tomography (CT), recurrent anatomy is predictable from context, whereas the sporadic, localized pathology that determines clinical interpretation often is not. We introduce TAP-JEPA, a ViT-B/3D trained on 113,611 CT volumes that is competitive with state-of-the-art encoders (COLIPRI, Merlin, VoxelFM) in scan-level classification. We then audit the learned predictive pathway by replaying the exact masking, teacher, and predictor computations used during training. Prediction lowers the mean Skill (prevalence-normalized average precision) from 0.388 to 0.271 for pathology, but only from 0.921 to 0.873 for anatomy. Along the teacher direction, separating pathology from matched control regions in the same organ, prediction moves pathology 34.7% toward the controls and controls 13.6% toward pathology, with larger shifts farther from visible context. Replacing visible matched-control representations increases surrounding-target prediction loss more than fully replacing pathology representations across four pathology categories. Downstream, removing all final-layer pathology tokens reduces the decision margin by at most 0.45% in four of seven scan-level probes. Together, these results show that contextual predictability is not a sufficient proxy for clinical relevance in medical representation learning. Training objectives should therefore retain diagnostic features regardless of their predictability from surrounding anatomy.

1 Introduction

JEPAs predict hidden-region teacher features from visible context, so their learning signal depends on what that context can resolve. In CT, this raises a concern that recurrent anatomy may be predictable while sparse pathology is not; the paper names this dependence predictability bias and audits it in TAP-JEPA.

1 Introduction

  • JEPAs learn visual representations by predicting teacher features at hidden locations from visible context, using structure inferred across partial observations.In natural images, spatial continuity makes much of a hidden region predictable from its surroundings.Source: §1 · p. 1
  • Figure 1 contrasts common masks that can hide localized pathology from the visible context with TAP-JEPA predictions that shift pathology representations toward matched normal controls.The caption points to Appendix C for the values.Source: Figure 1 · p. 2
  • CT anatomy recurs predictably, but sparse, heterogeneous pathology may be absent from anatomically compatible locations and remain unpredictable when hidden.This distinction matters as JEPA learning is being adopted across medical imaging.Source: §1 · p. 2
  • The paper calls the predictive objective’s dependence on what visible context can resolve predictability bias and studies anatomy versus pathology in TAP-JEPA.TAP-JEPA is a ViT-B/3D trained on 113,611 CT volumes and is competitive with state-of-the-art CT encoders on thoracic and abdominal classification benchmarks.Source: §1 · p. 2

2 Related Work

Prior work develops latent prediction across image, video, and medical-imaging settings, while studies of feature preference and diagnostic representation leave open whether the predictive objective itself removes local pathology. This paper connects these lines by tracing pathology through a 3D CT JEPA’s predictive computation.

2 Related Work

  • Latent prediction spans contextualized teacher features, spatial image blocks, and video, learning from structure shared between visible and hidden regions without reproducing every observation detail.The passage names Data2vec, I-JEPA, V-JEPA, and V-JEPA 2.Source: §2 · p. 2
  • Medical latent-prediction models cover radiographs, ultrasound, neuroimaging, and volumetric CT, alongside CT encoders trained with self-distillation or report supervision.Examples include CheXWorld, RadJEPA, US-JEPA, NeuroVFM, Neuro-JEPA, Rad-JEPA 3D, VoxelFM, Merlin, and COLIPRI.Source: §2 · pp. 2–3
  • Prior studies find objective-dependent support for global anatomy versus localized diagnostics and persistent difficulty with small, low-contrast findings, but do not trace findings through the training objective.Whether the objective itself removes local pathology remains open.Source: §2 · p. 3
  • Related theory studies predictable features, task-irrelevant variation, and context-unpredictable but task-relevant information; this work traces localized pathology through a large-scale 3D CT JEPA’s predictive computation.The passage connects these lines of work to studies of self-supervised medical representations.Source: §2 · p. 3

3 Predictability Bias

The section defines predictability bias as point-prediction learning’s dependence on what visible context resolves about the target. It characterizes the conditional prediction and training signal, then derives measurable implications for anatomy and pathology.

3 Predictability Bias

  • At each update, the student predicts hidden-target teacher features from visible context, while the teacher follows the student by exponential moving average.The student and predictor receive gradients through the predictions while teacher targets are fixed for that update.Source: §3 · p. 3
  • The student encodes visible inputs, the teacher encodes the complete view, and the predictor estimates teacher features at sampled target locations from context features and the mask.The prediction error is measured in teacher-feature space with a feature-wise L1 loss; target weights depend on body occupancy rather than clinical annotations.Source: §3 · p. 3
  • The best context-measurable point prediction is determined by the conditional target law; finite encoder and predictor capacity can add error.For L2 loss the optimum is the conditional mean, and for feature-wise L1 it is a coordinate-wise conditional median.Source: §3 · p. 4
  • A finding can be explicit in a hidden target yet unresolved from the predictor’s information, so its target-specific loss component contributes no conditional mean training signal through the visible path.The target-dependent score can still affect individual updates; under L2 it is proportional to the residual from the conditional mean.Source: §3 · p. 4
  • Predictability bias is point-prediction learning’s dependence on what visible context can resolve about the target.Optimal predictions summarize the conditional target distribution, while unresolved target-specific variation contributes no conditional mean training signal through the visible path.Source: §3 · p. 4
  • The framework predicts stronger preservation of context-constrained anatomy than sporadic pathology, contraction of pathology–control differences, and stronger attenuation when context is less informative.Visible-pathology interventions provide a context-side test by measuring their contribution to hidden-target prediction.Source: §3 · p. 4

4 Experimental Setup

TAP-JEPA is trained on a large multi-cohort thoraco-abdomino-pelvic CT dataset with volumetric masking and an occupancy-weighted feature loss. The evaluation replays the training prediction pathway and assesses frozen encoder features and regional discrimination using specified controls and metrics.

4 Experimental Setup

  • TAP-JEPA uses a ViT-B/3D student, an exponential-moving-average teacher, and a six-layer predictor, trained on 113,611 thoraco-abdomino-pelvic CT volumes.Training ran for 240 epochs; realized masks expose about 16% of in-body tokens as context and assign 84% as hidden targets.Source: §4 · pp. 4–5
  • Prediction analyses replay training tokenization, mask sampling, teacher, and predictor at a fixed checkpoint without updates; downstream evaluations use frozen encoder features from fully visible inputs.Regional analyses use unseen ATLAS and ReX scans, with annotation-negative controls and matched controls translated within the parent organ while preserving shape and voxel count.Source: §4 · p. 5
  • Encoder benchmarks report macro AUROC and AUPRC, while regional analyses use prevalence-normalized AP, Skill = (AP−π)/(1−π).Skill is zero at prevalence-level AP and one for perfect discrimination; uncertainty and multiplicity corrections accompany each analysis.Source: §4 · p. 5

5 Results

Hidden-target prediction preserves more teacher-accessible anatomical than pathological discrimination, and pathology contributes less to surrounding prediction than matched controls. Complete-view tokens show strong anatomical and positional organization, but scan-level dependence on local pathology varies.

5.2 Training masks frequently hide local pathology evidence

  • Across seven categories, 61.5–81.4% of pathology-mask events were fully hidden and targeted, while pathology tokens accounted for 0.41–7.04 per 1,000 target tokens.Median annotation-boundary distances to visible context were 20.25–25.11 mm.Source: §5.2 · p. 5

5.1 Competitive transfer across thoracic and abdominal CT

  • TAP-JEPA ranked highest on held-out RAD-ChestCT AUPRC (0.5268) and Merlin AUROC and AUPRC (0.8714 and 0.6965).Source: §5.1 · p. 5, Table 1 · p. 6

5.3 Target prediction reduces pathology discrimination more than anatomy

  • Pathology Skill shortfall exceeded anatomy in both cohorts: 36.7% vs 11.2% in ATLAS and 31.8% vs 6.2% in ReX.The paired pathology-minus-anatomy shortfall difference was 25.5 and 25.6 percentage points in ATLAS and ReX, respectively.Source: §5.3 · p. 6

5.4 Hidden-target prediction attenuates pathology–control separation

  • Attenuation ranged from 17.9% for scarring/fibrosis to 72.4% for liver lesions; atelectasis/consolidation showed 54.5% attenuation, with 48.2% pathology-side and 6.3% control-side displacement.Kidney lesions were the exception, with control-side displacement (28.2%) exceeding pathology-side displacement (15.4%); liver attenuation rose from 67.4% below 24 mm to 83.4% at 96 mm or more.Source: Figure 2 · p. 7, §5.4 · p. 7

5.5 Pathology context contributes less to surrounding prediction

  • Control replacement raised shared-target loss more than pathology replacement in all six categories; four differences remained significant after joint correction.In liver, scarring, atelectasis, and ground-glass, differences were +12.6% to +27.3% of control replacement cost; the nodule difference did not remain significant, and kidney results were descriptive.Source: §5.5 · p. 7, Table 2 · p. 8

5.6 Strong anatomical organization and uneven local pathology dependence

  • Complete-view tokens grouped by anatomical structure and showed smooth within-organ position gradients, while disease-state separation was less apparent.The UMAPs visualize encoder tokens without the predictor; disease-state separation varied with pathology token fill.Source: §5.6 · p. 8, Figure 3 · p. 9
  • Deleting pathology-overlapping tokens reduced margins by at most 0.45% for liver, kidney, scarring, and nodules, versus 10.47% for atelectasis and 2.39% for ground-glass.Alternative scan-level cues reached 0.68–1.24 times encoder Skill in five categories. The exploratory location-typicality association was ρpos = +0.085, positive in six of seven categories and nominally significant in three.Source: §5.6 · pp. 8–9, Table 3 · p. 10
  • The analyses characterize predictability bias in a large-scale 3D CT JEPA; objective changes and extensions to other partial-observation objectives remain future directions.Source: §5.6 · p. 9

6 Conclusion

TAP-JEPA captures strong anatomical and positional structure, but hidden-target prediction preserves substantially less of the signal carried by sporadic pathology. The conclusion argues that medical predictive pretraining should retain clinically important variation that surrounding anatomy cannot resolve.

6 Conclusion

  • Hidden-target prediction preserves substantially less signal from sporadic pathology than from anatomy in TAP-JEPA.The model is organized around what anatomy makes predictable, while clinically decisive deviations are less reliably retained.Source: §6 · p. 10
  • Pathology is shifted toward matched normal controls, contributes less than matched normal tissue to predicting surrounding anatomy, and can be removed with almost no decision change in several downstream tasks.The conclusion summarizes these findings as evidence that contextual predictability does not guarantee retention of clinically decisive deviations.Source: §6 · p. 10
  • The proposed next step is to preserve clinically important variation that surrounding anatomy cannot resolve while retaining anatomical structure useful for localization, shape, and spatial organization.The conclusion frames this as a direction for medical predictive pretraining.Source: §6 · p. 10

Declarations

The declarations describe the study's support, contributions, data sources, and use of existing de-identified research datasets. The appended propositions state that the pretraining criterion has no independent utility term and characterize optimal point predictions and the limits of deployed prediction residuals.

Declarations

  • The study used previously released, de-identified research datasets and reports funding, author contributions, and LLM-assisted drafting and editing.The funders had no role in study design, data collection, analysis, interpretation, or writing.Source: Declarations · pp. 10–11

A Importance Invariance

  • Reassigning downstream importance while preserving the law of pretraining variables leaves population risk and stochastic-gradient law unchanged.Clinical relevance can still influence learning through correlations in those variables; the invariance does not imply that more data cannot improve pathology representations.Source: A Importance Invariance · p. 16

B Optimal Predictions and Gradient Identities

  • Under squared error, the optimal point prediction is the conditional mean, while under absolute error it is a conditional median, which need not be unique.The result applies coordinate-wise to the vector target; the deployed feature-wise L1 scaling does not change its minimizers.Source: §B.1 · p. 16

B.1 Optimal predictions and the deployed loss

  • A deployed prediction residual does not isolate irreducible uncertainty because finite encoder or predictor capacity and optimization can add error.The optimum over predictions measurable from raw visible information may not be attainable by the deployed model.Source: §B.1 · p. 16

B.2 Proof and loss-specific forms of Lemma 1

The proof centers the chosen loss score conditional on visible information and gives distinct forms for squared and feature-wise absolute loss. It also limits what a zero conditional mean implies about updates during actual training.

B.2 Proof and loss-specific forms of Lemma 1

  • At a fixed step, the prediction and its parameter Jacobian are measurable from visible information, and subtracting the conditional mean gives the second identity in Eq. 4.The passage states this for the Jacobian where it exists.Source: §B.2 · p. 16
  • Under squared loss, the centered score is the negative doubled residual from the conditional-mean target.The unnormalized score is twice the prediction-minus-target difference, and centering yields −2(Z − E[Z | U]).Source: §B.2 · p. 16
  • For feature-wise L1 loss, the score depends on sign probabilities rather than only on the conditional mean of the target.The centered score is the chosen score minus its conditional expectation, not a centered target residual.Source: §B.2 · p. 16
  • A zero conditional mean for the centered score does not eliminate its contribution to individual updates in actual training.Training sums weighted target losses and updates shared parameters.Source: §B.2 · p. 17

C Estimands and Protocols

The paper distinguishes scan-level transfer, hidden-target recovery, matched-region displacement, visible-context contribution, and complete-volume readouts. It specifies the estimands and evaluation protocols, while reporting summary values as unweighted descriptive averages across datasets or pathology categories.

C Estimands and Protocols

  • The protocols separate scan-level transfer, pathology exposure, hidden-target recovery, matched-region displacement, visible-context contribution, and complete-volume readouts.Skill uses each task's eligible-label prevalence, and AP denotes stepwise average precision, including benchmark columns labelled AUPRC.Source: C Estimands and Protocols · p. 17
  • Pathology-side and control-side displacement quantify contributions to retained contrast along a cross-fitted pathology-sensitive teacher direction.The origin-constrained slope is retained contrast; matched isotropic axes calibrate generic attenuation of arbitrary feature directions.Source: C Estimands and Protocols · p. 17
  • Replacing visible pathology or matched-control activations measures direct predictive contribution through final context tokens on same-organ targets 24–96 mm away.The intervention occurs after student encoding, so information already propagated through self-attention remains available; an unrelated same-organ replacement calibrates the loss increase.Source: C Estimands and Protocols · p. 17
  • Abstract and figure headline values are unweighted means across datasets or pathology categories, not event-weighted pooled estimates, and have no separate uncertainty intervals.The Skill pairs average pooled ATLAS and ReX tasks; the displacement percentages average six pathology categories.Source: C Estimands and Protocols · p. 17

D Pretraining and Evaluation Details

  • The cohort counts stored volumes rather than unique patients, and evaluation exclusions and annotation support differ by task and cohort.ATLAS groups are cases or studies because patient IDs are unavailable; annotation-negative controls are not verified clinically normal.Source: D Pretraining and Evaluation Details · pp. 18–19

D.1 Frozen scan-level benchmark protocol

The protocols evaluate frozen scan-level representations on held-out cohorts and describe token-level anatomy and organ-state probes. Anatomy is decoded strongly, while acquisition metadata matches or exceeds encoder Skill for kidney and liver but not pancreas.

D.1 Frozen scan-level benchmark protocol

  • CT-RATE scan-level readouts were selected on validation data and evaluated on 3,002 quality-controlled volumes from the official validation split.The protocol crossed five token-aggregation families with four learning rates for TAP-JEPA and VoxelFM.Source: §D.1 · p. 19
  • RAD-ChestCT transfer applied CT-RATE-selected readouts unchanged, without fitting, selection, calibration, or thresholding on the external cohort.The COLIPRI-CRM comparison is literature-based rather than paired: its split details and test set differ from the local evaluation.Source: §D.1 · p. 20
  • The Merlin abdominal task used 15,174 fitting, 5,016 validation, and 5,082 test studies; patient-disjoint splitting relies on the publisher because identifiers were unavailable.This controls data and selection protocol across encoders, but not the exact readout head; discrimination does not establish calibration or clinical utility.Source: §D.1 · p. 20

D.2 Frozen anatomy, organ-state protocols

  • Anatomy-token probes achieved macro AUROC/Skill 0.9844–0.9852/0.7527–0.7610 for ATLAS and 0.9757–0.9839/0.7159–0.7334 for CT-RATE.Intervals condition on fitted probes and do not account for patient clustering of repeated CT-RATE scans.Source: §D.2 · p. 21
  • Acquisition matches or exceeds encoder Skill for kidney and liver, but not pancreas; neither organ-state decoding nor metadata-only probing localizes the signal to pathology.The anatomy probe uses final-layer foreground tokens and classifies annotated structures.Source: §D.2 · pp. 20–21

E Additional Frozen-Encoder Benchmark Results

The appendix separates source-domain frozen-encoder benchmarking from additional analyses of token organization, scan-level readout, localization, masking exposure, and teacher sensitivity. These analyses include explicit scope and measurement caveats for qualitative comparisons and derived estimates.

E Additional Frozen-Encoder Benchmark Results

F.1 Qualitative thoracic token representations

  • Pathology-bearing and report-normal lung tokens intermix in UMAPs, but cohort differences prevent separating disease-related structure from unmeasured differences.The visualization does not establish absence of pathology information or quantify disease discrimination.Source: §F.1 · p. 23

F.2 Scan-level MIL and alternative routes

  • Acquisition alone reaches 0.68–1.08 times encoder Skill across ATLAS classes; kidney encoder/acquisition Skill is 0.584/0.632, with 0.18% own-token necessity.ReX co-mention reaches 1.08 times encoder Skill for scarring and 1.24 for nodules; atelectasis/consolidation and ground-glass have necessities of 10.47% and 2.39%.Source: §F.2 · p. 24

F.3 Positional recurrence

  • Volume-adjusted typicality is positively associated with pathology-token ablation necessity (ρpos = 0.085 [0.042, 0.127]); its association with attention is 0.025 [−0.018, 0.067].The pooled intervals do not model dependence across categories sharing scans or patients, or uncertainty in estimated location maps.Source: §F.3 · p. 25

F.4 Attentive-MIL localization protocol

  • The localization protocol compares pathology-token attention with surrounding parent-organ or lobe tokens, using scan-level ratios and volume-stratified summaries.Table 10 also describes score–volume and global-attention-excess–volume correlations.Source: §F.4 · pp. 25–26, Table 10 · p. 26

F.5 Production-mask exposure

  • Production-mask exposure is measured over crop-retained pathology–mask events; same-organ distance summaries exclude events without visible same-organ context.For liver, the any-context median is 24.83 mm over 1,169,151 events, while the same-organ median is 24.50 mm over 1,033,029 events.Source: §F.5 · pp. 26–27, Figure 5 · p. 26, Table 11 · p. 27

F.6 Teacher-sensitivity gate

  • All six pathology families passed the teacher-sensitivity gate: at least 100 pairs, at least 10 of each role per fold, AUROC above 0.60, and bootstrap lower bound above 0.55.ReX uses positive–unlabelled evaluation, and its controls are unannotated rather than verified healthy.Source: §F.6 · p. 28, Table 12 · p. 28

F.7 Hidden-target recovery

Hidden-target recovery compares prediction, teacher, and a geometry-only floor at the same hidden targeted tokens. Prediction recovers more of the measured geometry-to-teacher Skill gap for anatomy than pathology in both datasets, with family-specific and sensitivity qualifications.

F.7 Hidden-target recovery

  • The analyzed cohorts comprise eligible annotated-source scans with at least one hidden-target pathology or control row, not all scans in either dataset.The prediction and normalized EMA-teacher target are read at the same hidden targeted token; concealment of a token does not imply concealment of the whole pathology.Source: §F.7 · pp. 28, 30
  • Paired anatomy-minus-pathology recovery differences are positive in both datasets; pancreas has a higher recovery ratio than kidney (0.779 versus 0.690) but lower prediction/teacher Skills (0.181/0.229 versus 0.427/0.580).The pooled results use class-macro Skill for anatomy and positive-class Skill for pathology.Source: §F.7 · p. 31, Table 15 · p. 32
  • Nodule/mass recovery is descriptive because its teacher–geometry gap is 0.073 [0.033, 0.130], below the 0.10 point-gap threshold.The limited teacher Skill, rather than a strong geometry floor, accounts for the small denominator; this family remains in pooled ReX results.Source: §F.7 · pp. 31–32
  • A richer geometry floor changes pooled anatomy/pathology recovery to 0.885/0.635 in ATLAS and 0.933/0.687 in ReX, while both paired anatomy advantages remain positive.Prediction and teacher scores are unchanged under this sensitivity analysis.Source: §F.7 · p. 32, Table 17 · p. 32
  • Including pancreas changes expanded ATLAS pooled pathology recovery from 0.637 [0.561, 0.693] in 595 scans to 0.633 [0.561, 0.686] in 756 scans.The paired change is −0.004 [−0.020, 0.013].Source: §F.7 · p. 32

F.8 Matched-region contrast, displacement and robustness

Matched-region replay finds attenuated pathology–control separation in all six categories, with pathology-side displacement larger than control-side displacement in five. The estimates show robustness across control and context-distance analyses, alongside category-specific differences.

F.8 Matched-region contrast, displacement and robustness

  • The analysis pairs pathology with exact-shape translated controls within the same organ or lung lobe and uses held-out-fold teacher directions for projection.Pancreatic lesions are excluded for insufficient eligible pairs; ReX controls are unannotated, not verified healthy tissue.Source: §F.8 · p. 33
  • Primary matched-region displacement estimates are reported with pointwise 95% group-bootstrap intervals, and signed displacement is measured in teacher-contrast units.Source: Table 19 · p. 34
  • All 72 common-mask coefficients and their pointwise upper confidence limits are below one, while control choice changes magnitudes and makes nodule asymmetry less stable.On the consensus axis, alternatives raise coefficients by 0.082–0.103 for liver, 0.059–0.068 for ground-glass and 0.122–0.145 for nodules.Source: §F.8 · p. 36, Table 24 · p. 36
  • All six farthest-distance coefficients are below their nearest-distance values, with five sequences decreasing monotonically; scarring rises slightly in the final bin.Scarring changes from 0.786 to 0.802 in the final distance bin.Source: §F.8 · p. 36, Table 25 · p. 37
  • Primary intervals condition on saved crops, masks, trained axes and checkpoint, while distance strata use role-specific events and changing eligible pairs.Bootstrap intervals do not include refitting or pretraining uncertainty; the distance analyses support a broad trend rather than identical event comparisons.Source: §F.8 · p. 36
  • Kidney and scarring/fibrosis retain more contrast than the generic direction, but both retained-contrast coefficients remain below one.Kidney has greater control-side than pathology-side displacement; the measurements indicate heterogeneity across categories.Source: §F.8 · p. 37

F.9 Visible-region interventions and robustness

Visible-region interventions show higher replacement costs for matched controls than pathology in several supported comparisons, including shared-target liver and scarring. The analysis also tests alternate replacements and target definitions, while its estimates depend on eligible events and fixed model state.

F.9 Visible-region interventions and robustness

  • Interventions replace selected final student activations while holding token identities, positions, masks, teacher targets and target weights fixed.The replacement uses a shared mean from clean donor tokens, and the adjusted contrast subtracts a count-matched reference calibration.Source: §F.9 · pp. 38–39
  • Corrected support remains for own-neighbourhood liver and atelectasis and for shared-target liver, scarring, atelectasis and ground glass.The joint correction spans both target definitions.Source: §F.9 · p. 40, Table 29 · p. 40
  • 12.8% of control replacement cost for shared-target liver and 27.3% for shared scarring are adjusted control-minus-pathology differences.For liver, the raw reduction is 11.9% and the adjusted difference is −0.60% of baseline local loss; scarring’s raw reduction is 25.0%.Source: §F.9 · p. 40
  • Liver and all four ReX categories pass the pre-intervention student-separation gate, but this diagnostic does not identify what the predictor uses.Replacement after encoding leaves information already propagated through student self-attention available in other tokens.Source: §F.9 · p. 41, Table 31 · p. 41
  • Empirical-donor replacement retains negative local contrasts with pointwise intervals below zero for liver, scarring, atelectasis and ground glass.The shared-target version also yields a negative nodule interval, but that comparison is auxiliary.Source: §F.9 · p. 41, Table 33 · p. 42
  • Partial replacement has positive point estimates while full replacement has negative point estimates, but the partial-dose intervals include zero for shared scarring and half-dose nodules.The covariate-adjusted intercept is descriptive: it relies on additive linear extrapolation despite nonlinear dose response and has no multiplicity correction.Source: §F.9 · pp. 42, 44, Table 35 · p. 43, Table 37 · p. 44

G Implementation checks and analysis limitations

The implementation checks replay the deployed forward path and document distinct draws and analysis-specific audit procedures. They also mark support boundaries for the visible-context intervention.

G Implementation checks and analysis limitations

  • A held-out replay reproduced image, occupancy, grid, masks, RoPE, features and loss without discrepancy against the deployed forward path.The exposure, feature-audit, recovery, displacement/intervention and exposure analyses use distinct draw configurations.Source: G Implementation checks and analysis… · p. 44
  • Mask-exposure analysis uses geometry without a model forward pass, distinguishing soft voxel-mass from occupied-token exposure.Its distances are voxel-boundary-to-visible-token-box distances, not token-centre distances.Source: G Implementation checks and analysis… · p. 44
  • Kidney is descriptive in visible-context interventions because it fails the support gate.Other intervention controls include own-neighbourhood and shared-target estimates, reference-set subtraction and joint reporting corrections.Source: G Implementation checks and analysis… · p. 45
Loading 2610.13237v1…