PaperlayerView exact revision on arXiv

Source-linked AI summary

ReVision3D: Attribution-Guided Recursive Self-Improvement for 3D Medical Perception

Ho Hin Lee, Yuyin Zhou, Yannan Yu, Shi Gu, Yifan Wu

arXiv:2609.32984v2cs.CV

TL;DR

Volumetric RSI lacks reliable feedback on where visual evidence is lost and which system component to revise. ReVision3D uses annotated volumes for spatially grounded replay, attribution, and targeted recursive revision. On abdominal CT, it achieved 0.79 liver-lesion recall and 0.83 kidney-tumor recall at under 0.4 false positives per patient.

  • Problem

    Existing RSI feedback may not identify where volumetric visual evidence is lost or whether localized findings are correct, leaving the intervention target unclear.

    Source: §1 · pp. 1–2

  • Method

    ReVision3D uses annotated CT volumes as replay worlds to attribute missed lesions across acquisition, perception, training, and inference, then evaluates targeted revisions under a fixed verifier and objective.

    Source: §1 · p. 2, §3 · p. 3

  • Results

    0.79 liver-lesion recall and 0.83 kidney-tumor recall at under 0.4 false positives per patient, with the largest gains on 1–5 mL lesions.

    Source: §6 · p. 9

  • Takeaways & Limitations

    Failure attribution and externally grounded verification are useful principles for recursive improvement of medical imaging systems.

    Source: §7 · p. 10

    The routing rule was formulated retrospectively, so prospective routing remains untested on a new organ or modality.

    Source: §6 · p. 9

Abstract

from arXiv · show

Recursive self-improvement (RSI) offers a promising path for overcoming the limited visual capability of current medical imaging agents. Yet applying RSI to volumetric imaging remains difficult: failures can arise from acquisition, perception, training recipe, or downstream inference, while self-generated feedback and logged trajectories provide little guidance on which component should change. We introduce ReVision3D, an RSI system that leverages 3D volumes with spatially grounded annotations to determine where visual evidence is lost and recursively improve the corresponding visual capability. A frozen language-model designer proposes revisions to acquisition, perception, training, or inference, while the verifier and system-level objective remain fixed. Our key insight is that an annotated volume forms an exact replay world for view rendering and spatial verification: unvisited views can be rendered on demand, and localized predictions can be checked directly against reference masks. This grounded feedback directs targeted revision, while only changes that improve beyond measured seed noise are retained. Each accepted change triggers renewed attribution, allowing the dominant bottleneck to shift across rounds. On abdominal CT, attribution identifies perception as the dominant remaining limitation. Revising that level enables ReVision3D to achieve 79% liver recall and 83% kidney recall at under 0.4 false positives per patient, outperforming the evaluated frozen multimodal foundation models, with the largest gains on small lesions. Our framework is public available with interactive demo in this project page: \url{https://leeh43.github.io/ReVision3D/}

1 INTRODUCTION

ReVision3D uses annotated 3D volumes for grounded replay and verification, then attributes failures to guide targeted visual-system revisions. On abdominal CT, perception adaptation improves liver and kidney recall after recipe-level progress saturates.

Introduction

  • Logged trajectories cannot reveal whether unvisited views would expose missed findings, and self-generated judgments cannot independently verify localization.Effective visual self-improvement therefore needs both attribution of where evidence is lost and external verification of proposed revisions.Source: §1 · pp. 1–2
  • Annotated scans provide an exact replay world for rendering unvisited views, measuring lesion coverage, and verifying localized predictions in shared 3D coordinates.This supplies counterfactual view evaluation and grounded perception feedback without relying on the reader to judge itself.Source: §1 · p. 2
  • Perception adaptation improves liver and kidney recall, including for small lesions, after recipe-level revisions stop moving the visual ceiling.Inference across successive readers also recovers complementary evidence, while the loop rejects changes within seed variability and exposes limitations outside its intervention space.Source: §1 · pp. 2–3
  • ReVision3D attributes missed lesions across acquisition, perception, training recipe, and inference, directing a frozen language-model designer to the corresponding intervention level.The verifier and system-level objective remain fixed; candidates are retained only when gains exceed measured variability, and accepted changes trigger renewed attribution.Source: §1 · p. 2

2 RELATED WORK

ReVision3D connects medical imaging agents with recursive self-improvement by revising visual reading components using counterfactual observations from annotated 3D volumes and a fixed verifier.

Medical imaging agents and grounded perception

  • ReVision3D treats acquisition, perception, training, and inference as separate intervention spaces, using volumetric replay to evaluate counterfactual views without reader inference.Related approaches extend multimodal models to CT with slice rendering and tools, train readers with scan-grounded rewards, or learn active view-acquisition policies.Source: §2 · p. 3

Recursive self-improvement and self-evolving agents

  • Unlike prior recursive-improvement approaches, ReVision3D revises the visual reading process using unlogged counterfactual observations from annotated 3D volumes.Prior work searches executable artifacts or evolves agent memory, skills, tools, and policies; ReVision3D uses volumetric replay and loss attribution to choose among four revision targets under a fixed verifier.Source: §2 · p. 3

3 REVISION3D: RECURSIVE SELF-IMPROVEMENT THROUGH VOLUMETRIC REPLAY

ReVision3D uses annotated CT volumes as replay worlds to locate where visual evidence is lost, then recursively revises the corresponding part of the medical reading system. Rendering, reference masks, verification, and the system-level promotion objective stay fixed, so candidate changes are judged against the same grounded criterion.

Volumetric replay

  • Annotated volumes enable counterfactual replay of where to look and grounded replay of what was localized, with deterministic rendering and spatially mapped predictions.Replay can render unvisited views, but it does not assume a reader’s response to them: the reader must actually run on those views.Source: §3 · pp. 3–4

Revision cycle

  • Each cycle attributes missed lesions, has a frozen language-model designer revise a corresponding component, and retains changes only under a fixed evaluator and objective.The verifier, reference masks, objective, and acceptance rule remain fixed; development patients determine promotion, and final multi-organ evaluation occurs only after choices are frozen.Source: §3 · pp. 3–4

Grounded feedback

  • Using a shared 5 mm tolerance, replay distinguishes whether a view geometrically exposes a lesion from whether a reported location matches a reference lesion.Exposure can be evaluated from scan geometry and masks without invoking the reader; grounded matching yields hits, duplicates, and ungrounded findings, while bounded training-reward weights can shift the precision-recall trade-off without redefining correct localization.Source: §3 · p. 4

Failure attribution

  • Replay decomposes missed-lesion losses across pipeline stages, but finite sampled reads cannot establish that a lesion is truly unreachable, and attribution is not causal.On development and test patients, the two middle terms are reported jointly as the never-detected share.Source: §3 · p. 4

Revision levels

  • Four loss types route revisions to acquisition, perception, training recipe, or inference; a learnability census heuristically directs undetected losses toward training or perception.Acquisition candidates are screened by geometric exposure; perception candidates require actual training; training uses groups with non-identical sampled rewards; inference scores merged 3D candidates using label-free features.Source: §3 · pp. 4–5

Promotion and recursion

  • Replay screens rank but do not promote candidates; accepted changes must pass level-specific tests and trigger renewed attribution, while unsuccessful search can exhaust a level.Training requires improvement beyond the incumbent seed range and positive gains for every matched seed; inference requires at least 0.9 paired-bootstrap probability of improvement. The routing rule was formulated retrospectively and did not schedule the reported experiments.Source: §3 · p. 5

4 EXPERIMENTAL DESIGN

ReVision3D is evaluated for liver-lesion and kidney-tumor localization on abdominal CT, using spatially verified predictions and a fixed development-to-test protocol. The design compares uniform-slice frozen-model baselines with dense-read variants and evaluates recall across lesion sizes and false-positive operating points.

4 EXPERIMENTAL DESIGN

  • LiTS and KiTS23 evaluate liver lesions and kidney tumors; a hit lies within 5 mm of a reference mask, and the primary analysis includes lesions ≥1 mL.The reader is Qwen3-VL-4B-Instruct and the default designer is Claude Opus 5; choices are made on development patients before final multi-organ evaluation.Source: §4 · p. 5
  • Frozen frontier and open models read eight uniform axial slices, while ReVision3D processes axial slices every 5 mm and merges neighboring findings into 3D candidates.Candidate support is the number of slices reporting it; varying the support threshold traces a recall–false-positive curve. The study also compares Gemini 3.8 Flash and reader variants under the same dense-read protocol.Source: §4 · p. 5
  • Survey-baseline recall is reported at ≤0.5 false positives per patient overall and across four lesion-volume bins; revisions are promoted using J from Eq. (2).The bins are 1–5, 5–20, 20–100, and >100 mL.Source: §4 · p. 5
  • Figure 2 compares frozen frontier models with ReVision3D on 31 liver test patients and 111 lesions ≥1 mL, including recall-versus-false-positive and lesion-volume views.Its volume-bin differences use Gemini 3.8 Flash reading every slice under the study protocol as the comparator; Table 1 reports recall by lesion volume for liver and kidney test patients.Source: Figure 2 · p. 6, Table 1 · p. 6
  • Confidence intervals use patient-level bootstrap resampling, and paired systems are compared with paired bootstraps.Source: §4 · p. 6

5 RESULTS

On held-out abdominal CT, ReVision3D outperformed the evaluated frozen readers, with especially large gains on small lesions. Recursive attribution and ablations identify perception adaptation—not additional slices or recipe changes alone—as the main source of improvement.

5.1 COMPARISON WITH FRONTIER MODELS

  • At ≤0.5 false positives per patient, ReVision3D achieved 0.79 liver recall versus 0.05–0.36 across thirteen frozen readers, and 0.83 kidney recall versus 0.70 for the strongest frozen reader.For 1–5 mL lesions, recall was 0.69 on liver and 0.57 on kidney; no frozen reader exceeded 0.13 and 0.21, respectively.Source: §5.1 · pp. 6–7

5.2 RECURSIVE IMPROVEMENT TRAJECTORY

  • Recipe revisions raised liver J from 0.379 to 0.529, but none of fourteen further proposals exceeded measured noise; attribution then assigned the dominant remaining loss to perception.With the frozen encoder, never-detected lesions accounted for 0.47 of liver and 0.40 of kidney test lesions; encoder adaptation raised reachable recall from 0.57 to 0.70, and a subsequent inference revision reached J = 0.608.Source: §5.2 · p. 7

5.3 PERCEPTION LEVEL AND TEST EVALUATION

  • After recipe-level progress saturated, adapting more vision blocks raised mean liver-kidney objective ¯J from 0.381 to 0.535, while a two-epoch configuration reached ¯J = 0.716 and was accepted.The accepted configuration reduced the never-detected share to 0.14 on liver and 0.26 on kidney in its first trial.Source: §5.3 · p. 8, Figure 4 · p. 8
  • On held-out tests, designer-selected perception achieved mean J = 0.75 and 0.66 across seeds, versus 0.54 for human-selected adapters and 0.43 for a frozen encoder; selected inference reached liver/kidney J = 0.82/0.83.Inference-level search found an accepted program in all twelve runs across four designers from three model families.Source: §5.3 · pp. 8–9, Table 2 · p. 8

5.4 ABLATION ANALYSIS

  • Ablations found that perception configuration produced the largest gain, while training signal, reading protocol, and recipe changes yielded more limited improvements.A supervised warm start raised survey recall from 0.05 to 0.25, while sparse reinforcement learning reached 0.27; dense reading raised overall recall from 0.27 to 0.45 and second-reader confirmation to 0.56. Recipe changes kept overall recall within 0.41–0.49, whereas 1–5 mL recall under the same dense protocol rose from 0.06 with a frozen encoder and 0.12 with four hand-selected blocks to 0.69 with the designer-selected configuration.Source: §5.4 · p. 9

6 DISCUSSION

ReVision3D improves medical visual capability through external feedback, with perception adaptation reducing never-detected findings and raising the reachable recall ceiling. Its discussion also identifies untested routing and rendering choices as limitations.

6 DISCUSSION

  • 0.79 liver-lesion recall and 0.83 kidney-tumor recall at under 0.4 false positives per patient were achieved, with the largest gains on 1–5 mL lesions.These are the final reader's reported results.Source: §6 · p. 9
  • Similar inference-level outcomes across four tested designers suggest that fixed external evaluation reduces dependence on the model proposing revisions.Source: §6 · p. 9
  • Recipe revisions refine existing detections, dense reading exposes more findings, and perception adaptation reduces the never-detected share to raise the reachable recall ceiling; re-attribution shifts search as bottlenecks move.The tested interventions affect different failure modes and are not interchangeable.Source: §6 · p. 9
  • Prospective routing remains untested on a new organ or modality, and the current loop fixes rendering choices such as windowing and neighboring-slice context.The routing rule uses training-patient learnability to estimate which intervention may reduce development loss and was formulated retrospectively; extending the intervention space to rendering variables is a natural next step.Source: §6 · p. 9

7 CONCLUSION

ReVision3D uses annotated CT volumes as replay environments to attribute lost visual evidence and guide evidence-gated recursive improvement. On liver and kidney CT, perception was the dominant remaining limitation, with its revision producing the largest gains, particularly for small lesions.

7 CONCLUSION

  • Annotated CT volumes serve as replay environments where unvisited views can be rendered, localized findings checked against fixed geometry, and missed lesions attributed to where evidence is lost.Source: §7 · p. 9
  • The framework applies an evidence-gated improvement loop across acquisition, perception, training, and inference.Source: §7 · p. 9
  • On liver and kidney CT, perception was the dominant remaining limitation and its revision produced the largest gains, particularly for small lesions, while unsupported revisions were rejected.The conclusion presents failure attribution and externally grounded verification as useful principles for recursive improvement of medical imaging systems.Source: §7 · pp. 9–10

AI USE STATEMENT · A TERMINOLOGY USED IN TABLES

The paper states how AI tools contributed to the method and editing, and defines recurring terms used to interpret its tables. These statements also specify the work’s ethical scope and reproducibility materials.

AI USE STATEMENT

  • A frozen designer proposes system revisions, thirteen frontier models serve as baselines, and authors reviewed and verified AI-assisted language edits and table and figure scripts.The designer models named are Claude Opus 5, GPT-6 Astra, Gemini 3.1 Pro, and Gemini 3.8 Flash in Appendix G; the authors take responsibility for the paper’s full content.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

  • The work uses public, de-identified LiTS and KiTS23 CT data, collects no new patient data, runs no human-subject experiments, and presents ReVision3D as non-clinical lesion-localization research.Reference masks are used for training, replay, and evaluation; the system makes no diagnosis claim, defers unverifiable clinical judgments to clinicians, and benchmark performance may not transfer to other scanners, protocols, or patient populations.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

  • Section 3 defines the reward, objective, attribution identity, screening, and acceptance rule; Appendix B reports experimental settings and both accepted and rejected trials.Code, designer prompts, playbooks, and trial registries are prepared for release with supplementary material.Source: REPRODUCIBILITY STATEMENT · p. 10

A TERMINOLOGY USED IN TABLES

  • The 8-slice survey presents eight evenly spaced axial slices in one prompt for frozen-model comparisons, while dense read evaluates every axial slice at 5 mm spacing for trained readers.Dense-read findings on neighbouring slices are merged into 3D candidates.Source: A TERMINOLOGY USED IN TABLES · p. 12
  • Lesion-guided slices intersect every annotated lesion to estimate oracle view-selection benefit, while the acquisition program starts with four survey slices and selects up to four additional views.Additional views are selected by the learned view-selection program ψ.Source: A TERMINOLOGY USED IN TABLES · p. 12
  • Support counts rendered slices reporting a candidate, and a support threshold of at least k retains candidates observed on k or more slices; increasing k typically lowers both false positives and recall.A cascade requires a second reader to confirm each candidate within 12 mm in shared 3D coordinates.Source: A TERMINOLOGY USED IN TABLES · p. 12
  • An operating point fixes an inference threshold to produce a recall–false-positive pair; R is lesion-level recall for lesions ≥1 mL, FP is false positives per patient, and J is the system-level objective.“Recall at ≤0.5 FP/patient” means the highest recall attainable at or below 0.5 false positives per patient.Source: A TERMINOLOGY USED IN TABLES · p. 12
  • Failure attribution distinguishes lesions never detected, detected but discarded during aggregation or thresholding, and detected and kept in the final prediction.These categories correspond to the decomposition in Eq. (3).Source: A TERMINOLOGY USED IN TABLES · p. 12

B EXPERIMENTAL DETAILS · C FROZEN FRONTIER READERS

The experiments use patient-disjoint LiTS and KiTS23 cohorts, compare multiple reading protocols and frozen frontier readers, and evaluate performance with lesion-level recall and patient-level false positives. The self-improvement procedure screens candidate revisions against seed noise before retraining leading proposals for acceptance.

B EXPERIMENTAL DETAILS

  • LiTS and KiTS23 provide 131 liver and 292 kidney CT volumes, with patient-disjoint splits for the liver and multi-organ studies.A fixed 30-patient development subset scores short trials, while the full 60-patient development set governs promotion.Source: B EXPERIMENTAL DETAILS · p. 13
  • The frozen reader is Qwen3-VL-4B-Instruct, with low-rank adapters trained for language-model and, in perception configurations, visual components.Volumes are reoriented to RAS, cropped to an organ box expanded by 15 mm, and rendered as axial slices; deployment uses an external organ segmenter to supply the box.Source: B EXPERIMENTAL DETAILS · p. 13
  • The reading protocols compare four- or eight-slice surveys, budgeted additional views, dense 5 mm reads, and lesion-guided oracle slices.The four-slice survey exposes only 83% of liver lesions; a slice within 5 mm of a small lesion may show only its edge.Source: B EXPERIMENTAL DETAILS · p. 13

C FROZEN FRONTIER READERS

  • Thirteen frozen frontier readers are evaluated, with reasoning models allowed up to 64k output tokens and one low-effort retry for empty or unparsable responses.The readers are Mistral Large 3, Qwen3.5-397B, Gemini 3.1 Pro, Gemma 4 31B, GPT-5.5, Claude Opus 5, Qwen3.8-Max, Qwen3.8-27B, GLM-5.3 Flash, Qwen3.8-Flash, Gemini 3.8 Flash, GPT-6 Astra, and DeepSeek V4.1 Flash.Source: B EXPERIMENTAL DETAILS · p. 13

B EXPERIMENTAL DETAILS

  • Recall is pooled over lesions ≥1 mL, false positives are counted per patient, and dense-read J is evaluated across operating points.Uncertainty uses a patient-level bootstrap with 2,000 resamples; false-positive differences use a paired Wilcoxon signed-rank test.Source: B EXPERIMENTAL DETAILS · p. 13
  • Recipe searches evaluate 4–10 proposals for 100 steps after 2–3 seeds establish the noise floor, then retrain leading proposals for 400 steps before acceptance.Perception trials use six single-change proposals on 120 patients and replicate the two best; inference programs are revised up to eight times on cached development reads. Acquisition is reported separately and does not pass the full promotion gate.Source: B EXPERIMENTAL DETAILS · p. 13

C.1 DOES BETTER ACQUISITION HELP FROZEN READERS? … E.4 LEARNABILITY CENSUS AND EXHAUSTION OF RECIPE-LEVEL SIGNAL

Across these experiments, additional views and recipe-level changes do not resolve the remaining detection failures; perception adaptation improves lesion detection, especially for small lesions, and attribution distinguishes newly detected findings from aggregation effects.

C.1 DOES BETTER ACQUISITION HELP FROZEN READERS?

  • Frozen readers gain little from better views, while small lesions expose the largest perception gap.Across 13 frozen frontier readers, the acquisition program improves the strongest reader by at most 0.10 and oracle lesion-guided slices by only 0.11; Gemini 3.8 Flash moves from 0.36 to 0.38 with acquisition. Below 5 mL, no frontier model detects more than 0.13 of liver lesions (7 of 52), whereas ten of thirteen detect at least 0.55 above 100 mL. For kidney tumors below 5 mL, no frontier model detects more than 3 of 14 (0.21), while ReVision3D detects 8 (0.57).Source: §C.1 · p. 14, Table A1 · p. 14, §C.2 · p. 15

C.3 SEPARATING READING PROTOCOL FROM PERCEPTION

  • 0.79 liver recall and 0.83 kidney recall follow designer-selected perception adaptation under the same dense protocol.Under that protocol, liver recall rises from 0.19 with a frozen encoder to 0.25 with hand-selected adapters and 0.79 with the designer-selected configuration; kidney recall rises from 0.60 to 0.70 and 0.83. For Gemini 3.8 Flash, dense reading alone leaves liver recall essentially unchanged at 0.36 to 0.34 and lowers kidney recall from 0.60 to 0.44.Source: §C.3 · p. 16, Table A3 · p. 16

D.1 ATTRIBUTION ACROSS PERCEPTION CONFIGURATIONS

  • Perception adaptation reduces never-detected lesions and, in the selected configuration, also reduces liver lesions discarded during aggregation.On liver, the never-detected share falls from 0.42 to 0.30 to 0.25 across frozen, hand-selected, and designer-selected configurations, while kept share rises from 0.45 to 0.50 to 0.70; the discarded share falls to 0.05 with the designer-selected configuration. On kidney, no detected tumor is discarded, and the never-detected share falls from 0.45 to 0.30 to 0.14. In the development cohort, designer-selected perception performs best in every sub-100 mL bin on both organs; test-set counterparts are reported in Table 1.Source: Table A6 · p. 17, §D.1 · p. 18, §D.2 · p. 18

D.3 SEARCHING THE PERCEPTION LEVEL

  • Adapting 12 and then all 24 vision blocks raises trial-scale ¯J by 0.115 and 0.155, while higher learning rates cause divergence.Increasing image resolution alone produces a gain equal to the 0.023 noise floor; a second training epoch on stronger adaptation yields the largest trial-scale objective, ¯J = 0.716.Source: Table A8 · p. 18, §D.3 · p. 19

E.1 ACCEPTED REVISIONS CHANGE DIFFERENT PARTS OF THE OPERATING CURVE

  • Across accepted revisions, J rises from 0.379 to 0.608, but only perception adaptation raises maximum liver recall from 0.57 to 0.70.Early training stages increase recall at 0.25 false positives per patient from 0.17 to 0.42 while maximum recall remains approximately 0.57–0.58; precision stays at least 0.88 throughout.Source: §E.1 · p. 19, Table A9 · p. 19

E.2 FIRST RECIPE SEARCH: SHORT-RUN GAINS DO NOT SURVIVE PROMOTION

  • Three short-run recipe trials clear the threshold, but the strongest candidates fall to +0.046 and +0.017 after 400 steps and are not promoted.The replay census overestimates realized learning-signal steps and ranks short-run proposals nearly uninformatively (Spearman ρ = 0.02); the hand-designed learnability recipe reaches J = 0.529 after full training as a reference. Revisions are retained only when fixed-objective improvement exceeds measured seed noise.Source: §E.2 · p. 20, Table A10 · p. 20, Figure A1 · p. 20

E.3 RECIPE SEARCH AFTER PERCEPTION ADAPTATION: NO FURTHER GAIN

  • After perception adaptation, no recipe proposal improves on J = 0.575 by more than the 0.042 noise floor, so the round is rejected.The strongest proposal reaches J = 0.592, only +0.017 over the incumbent, and its never-detected share remains within seed-run variation; attribution is recomputed before reconsidering the recipe level.Source: §E.3 · p. 21, Table A11 · p. 21

E.4 LEARNABILITY CENSUS AND EXHAUSTION OF RECIPE-LEVEL SIGNAL

  • After training on learnable fresh views, the frontier contracts to 7% overall, but 62% of lesion-bearing prompts remain never hit.Initially, 25% of cached-view prompts and 23% of fresh-view prompts lie on the learning-signal frontier, while 39% and 50% of lesion-bearing prompts are never hit. The census indicates recipe optimization saturates as reward-varying prompts are consumed and never-hit prompts provide no group-relative signal.Source: §E.4 · pp. 21–22, Table A12 · p. 21, Figure A2 · p. 21

E.5 THE TRAINING PROCEDURE TRANSFERS BETTER THAN THE READER … G.1 THE IMPROVEMENT LOOP IS ROBUST ACROSS TESTED DESIGNERS

The experiments show that ReVision3D’s training procedure transfers across organs better than a trained reader, while successive readers offer complementary evidence and inference-level improvement remains stable across tested designers. Successive recipe revisions do not raise the single-reader recall ceiling, and the available learning signal contracts.

E.5 THE TRAINING PROCEDURE TRANSFERS BETTER THAN THE READER

  • The liver-trained reader reached 0.16 kidney recall at 0.37 false positives per patient, while reusing the training procedure reached 0.58 recall at 0.33 false positives per patient after RL.The reused procedure reached 0.48 recall after supervised initialization; adding lesion-guided slices raised recall only from 0.58 to 0.61.Source: Figure A3 · p. 22, Table A13 · p. 22, §E.5 · p. 22

F.1 SUCCESSIVE RECIPE REVISIONS REACH A SIMILAR CEILING

  • Successive recipe-trained readers had similar dense-read recall, remaining within 0.41–0.49 with overlapping confidence intervals, indicating a largely unchanged single-reader detection ceiling.All results used dense reading with support ≥2 on 31 held-out liver patients.Source: Table A14 · p. 23, Figure A4 · p. 23, §F.1 · p. 23
  • Recipe revisions shifted precision and progressively exhausted learning signal: cached-view precision was 0.96 at 0.06 false positives per patient, while failure-mined precision was 0.69 at 0.74.Zero-variance steps rose to 53–88% for post-RL readers from 19% for the RL reader, while the learnable frontier shrank from 25% to 7% of candidate prompts.Source: §F.1 · p. 23, Table A14 · p. 23, Figure A4 · p. 23

F.2 SUCCESSIVE READERS PROVIDE COMPLEMENTARY EVIDENCE

  • Confirming the RL reader with the cached-view reader increased recall by 0.11 [0.03, 0.18] without a significant false-positive increase (p = 0.82).Fresh-view confirmation gained 0.18–0.20 recall at the cost of 0.52–1.26 additional false positives per patient.Source: Table A15 · p. 24, Figure A5 · p. 24, §F.2 · p. 24
  • Among individual readers, only the fresh-view reader significantly improved over the equal-compute control (+0.08 [0.01, 0.14]) and cached-view reader (+0.08 [0.04, 0.12]).Successive readers are more useful as complementary evidence than replacements, motivating inference that combines findings across readers rather than selecting one checkpoint.Source: §F.2 · pp. 24–25, Table A15 · p. 24, Figure A5 · p. 24

G.1 THE IMPROVEMENT LOOP IS ROBUST ACROSS TESTED DESIGNERS

  • All twelve inference-level searches across four designers and three incumbents produced an accepted program, with eleven accepted on the first revision.Within each incumbent setting, accepted objectives differed across designers by at most 0.005, 0.004, and 0.011, respectively; Gemini 3.8 Flash performed comparably as a designer despite weaker medical image-reading performance under the same protocol. The fixed external objective determines which proposed candidates are retained.Source: Table A16 · p. 25, §G.1 · p. 25

G.2 ABLATIONS LOCALIZE THE LARGEST GAIN TO PERCEPTION

On 31 held-out liver patients, component-wise ablations show that changing perception configuration produces the largest recall gains, especially for small lesions. Dense reading also improves small-lesion recall, while recipe variants remain similar.

Perception configuration

  • Perception changes raise 1–5 mL recall to 0.69 and overall recall to 0.79, versus 0.06 and 0.19 with the frozen encoder and 0.12 and 0.25 with hand-selected adapters.The ablations change one component at a time on the same 31 held-out liver patients.Source: §G.2 · p. 26, Table A17 · p. 26

Reading protocol

  • Dense reading raises 1–5 mL recall from 0.08 to 0.29, and confirmation by a second reader reaches 0.37.The comparison uses the dense protocol at the indicated false-positive constraint or support level.Source: §G.2 · p. 26, Table A17 · p. 26

Training signal

  • A supervised warm start raises overall survey recall from 0.05 to 0.25, mainly for lesions above 20 mL; RL and a 10-epoch supervised control leave 1–5 mL recall at 0.08–0.12.Block (a) uses the 8-slice survey.Source: §G.2 · p. 26, Table A17 · p. 26

Training recipe and attribution

  • Recipe variants remain within 0.23–0.31 on the 1–5 mL bin, while attribution indicates perception adaptation most increases the set of lesions detected.Acquisition and inference recover or reconcile evidence already within reach, whereas recipe changes primarily alter how that evidence is used.Source: §G.2 · p. 26, Table A17 · p. 26

H EXTENDED LIMITATIONS AND FUTURE DIRECTIONS

ReVision3D improves verifiable spatial perception, but its scope is bounded by the verifier, intervention space, replay capabilities, attribution method, and fixed objective. The passages identify extensions to broaden verification, interventions, and evaluation while retaining external checks.

Dependence on an external verifier

  • Reference lesion masks verify spatial localization, not characterization, malignancy, etiology, or other clinical judgments, so ReVision3D does not improve clinical reasoning as a whole.Future verifiers could incorporate longitudinal follow-up, pathology, structured measurements, or cross-modality agreement.Source: H EXTENDED LIMITATIONS AND FUTURE… · p. 26

Bounded intervention space

  • The designer searches only four predefined levels, limiting discovery of new backbones, modalities, rendering operators, or reasoning strategies unless exposed as actions.Future systems could expand the intervention space by proposing new visual modules, rendering strategies, or tool interfaces, subject to the same external promotion test.Source: H EXTENDED LIMITATIONS AND FUTURE… · p. 27

Replay is exact for geometry, not perception

  • Volumetric replay exactly renders unvisited views and measures geometric lesion exposure, but perception changes still require model runs, training, and evaluation.Lightweight surrogates could predict perception changes likely to reduce the never-detected share, with the fixed evaluator retained for final promotion.Source: H EXTENDED LIMITATIONS AND FUTURE… · p. 27

Attribution guides intervention but is not causal

  • The missed-lesion decomposition is exact, but finite sampled reads and the learnability census make training-versus-perception attribution a routing heuristic, not a causal guarantee.Future routing could estimate expected improvement from prior revisions or allocate a small exploration budget across levels before expensive training.Source: H EXTENDED LIMITATIONS AND FUTURE… · p. 27

A fixed objective defines the boundary of improvement

  • A fixed verifier and objective prevent the designer from redefining success, but localization gains may not improve downstream clinical decisions.The proposed next step is a hierarchical objective that retains spatial perception as a grounded inner loop and evaluates clinical utility separately with independent evidence and human oversight.Source: H EXTENDED LIMITATIONS AND FUTURE… · p. 27
Loading 2609.32984v2…