acceptodds
Under review as a conference paper at ICLR 2027

Cross-modal Heterogeneous Privileged Policy Optimization for Medical Image Understanding

Abstract

Foundation models for medical imaging now span self-supervised vision, dense supervision, and vision–language learning, offering complementary knowledge about structure, localization, and semantics. Clinical deployment, however, still favors compact vision-only models, and transferring from such heterogeneous teachers jointly is difficult because their modalities, objectives, and representation spaces do not match. We introduce HiPPO, an on-policy distillation framework that transfers heterogeneous multimodal foundation models into a compact vision-only student. Conditioned on the current student state, HiPPO samples candidate representations as binary selections over a unified codebook, scores each in the native space of every applicable teacher via teacher-specific projectors and compatibility functions, and reinforces candidates with stronger group-relative feedback. The selected representations drive position-wise visual distillation from frozen visual foundation models and privileged semantic distillation from a frozen medical vision–language model conditioned on annotation-derived views available only during training. We further establish MedBridge-Bench, spanning case-level disease diagnosis and voxel-level lesion recognition over more than 18,000 volumetric CT, MRI, and PET/CT studies. Across these settings, HiPPO improves compact medical vision models and attains competitive performance, while all teachers, privileged inputs, and policy modules are discarded at inference.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
medical vision, on-policy distillation, multimodal heterogeneous foundation model, group-relative advantage

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_8fYAOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

HiPPO distills knowledge from heterogeneous foundation models into a compact, vision-only 3D student network. Guided by local student features, a policy samples components from a shared codebook to construct candidate representations, which are then scored by compatibility rewards computed in each visual teacher’s native representation space. The policy is optimized with group-relative advantages, while the difficulty of matching each candidate modulates the spatial distillation weights. In addition, a separate branch provides privileged semantic supervision using lesion-centered views available only during training. We further introduce MedBridge-Bench, which evaluates classification and segmentation across CT, MRI, and PET/CT. At inference time, only the compact student is retained.

Critical strengths

The problem addressed in this paper is practically relevant, and the evaluation is reasonably broad. Transferring knowledge from multiple training-time teachers into a vision-only student—without requiring additional deployment inputs or auxiliary modules—is a meaningful and well-motivated objective. The authors also report training overhead, which is helpful for assessing the practical cost of the approach. The paper goes beyond reporting final task performance and provides supporting evidence at the mechanism level. The ablations—covering deterministic projection, soft mixtures, random sampling, policy removal, and the spatial weighting scheme—along with feature probes and gradient-path analysis, together offer a solid basis for understanding why and how the method works.

Critical weaknesses

The novelty claim does not sufficiently engage with the most directly related work on visual multi-teacher distillation. AM-RADIO [1], Theia [2], and DUNE [3] already combine teachers with heterogeneous objectives and representation spaces via teacher-specific mappings, and MTKD-RL [4] has explored policy-driven multi-teacher distillation for visual recognition. Consequently, heterogeneous teacher fusion, native-space alignment, or the use of a policy, taken individually, cannot be the sole basis for the paper’s main novelty. To be fair, HiPPO does offer a genuine distinction in that it operates on local codebook compositions rather than on teacher weights. However, the paper neither clearly articulates nor empirically validates what specific problem this change in the action space solves. Ablations over simple variants, while informative, cannot substitute for head-to-head comparisons against these directly related methods. [1] Ranzinger M, Heinrich G, Kautz J, et al. Am-radio: Agglomerative vision foundation model reduce all domains into one[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024: 12490-12500. [2] Shang J, Schmeckpeper K, May B B, et al. Theia: Distilling diverse vision foundation models for robot learning[J]. arXiv preprint arXiv:2407.20179, 2024. [3] Sarıyıldız M B, Weinzaepfel P, Lucas T, et al. Dune: Distilling a universal encoder from heterogeneous 2d and 3d teachers[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025: 30084-30094. [4] Yang C, Yu X, Yang H, et al. Multi-teacher knowledge distillation with reinforcement learning for visual recognition[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2025, 39(9): 9148-9156.

The attribution of the core mechanism, and of the external-generalization gains in particular, remains ambiguous. The sampled representations are trained with teacher rewards, whereas the deterministic student learns primarily through a separate weighted feature loss; as such, the contribution of discrete representation exploration needs to be disentangled from that of multi-teacher supervision and spatial hard-example weighting. In Table 5, the AUC improvement over deterministic projection is only 0.64 percentage points, and no evidence from repeated training runs is provided to establish whether this difference is statistically meaningful. Furthermore, the external kidney cohort mixes positive public cancer cases with private non-contrast negative controls, which risks confounding disease status with data source or acquisition protocol. Finally, the kidney sensitivity and specificity in Table 8 imply a BAcc of approximately 88.88%, whereas Table 2 reports 90.52%; if different thresholds or sample subsets were used for the two tables, this should be explicitly stated.

Questions for authors

Under matched teachers, students, supervision, and budgets, how does HiPPO compare with AM-RADIO/Theia-style feature distillation and policy-driven teacher weighting? Which results specifically favor representation-composition actions over teacher-weight actions?

What if codebook sampling is removed and difficulty weights are derived directly from deterministic student–teacher distances, with all other terms unchanged? This would test whether sampling adds information beyond direct hard-example weighting.

Please provide repeated-seed results for the full method, deterministic projection, and soft mixture. Also, what happens when the semantic branch’s soft tokens are removed or shuffled?

Could the external evaluation include matched positive and negative cases from the same sources and acquisition protocols? Please also reconcile the kidney metrics in Tables 2 and 8 and specify the threshold-selection partition.

7 comments
AuthorsOpenReview ↗

Relationship to Visual Multi-Teacher Distillation and the Contribution of HiPPO (1)

We thank the reviewer for this important pointer. AM-RADIO, Theia, DUNE and MTKD-RL are indeed the directly related work on the teacher side, and we will add them to Related Work in the revision. Our Related Work was organized around the two lenses through which we developed the method, on-policy distillation and privileged information distillation, and both are student-side lenses, so the agglomerative multi-teacher literature was never surfaced as a separate axis. We agree this is a gap in exposition and that expert-side work belongs in the discussion.

(1) Shared Foundations and Methodological Differences

Teacher-specific mappings are established components of multi-teacher distillation. AM-RADIO computes separate teacher-matching losses through teacher-specific heads and aggregates those scalar losses. Theia uses teacher-specific translators, while DUNE uses Transformer projectors to transfer heterogeneous global and patch-level knowledge. MTKD-RL learns sample-wise teacher weights for logit and feature distillation. HiPPO retains teacher-specific alignment but changes the object and spatial granularity of policy control.

DimensionAM-RADIOTheiaDUNEMTKD-RLHiPPO
Primary settingMulti-teacher consolidation into an image encoderVisual-encoder distillation for robot learningImage-encoder distillation from heterogeneous 2D and 3D teachersPolicy-driven distillation for visual recognitionHeterogeneous knowledge transfer into a 3D medical student
Teacher-specific alignmentMLP heads and separate teacher-matching lossesCNN/linear translators and separate feature lossesTransformer projectors for global and patch featuresLogit and feature distillationProjectors evaluate local codebook candidates and align deterministic student features
Policy actionNo distillation policyRobot policies are learned downstream of visual distillationNo local codebook-composition policySample-wise teacher weightsBinary codebook compositions conditioned on local student states
Supervision allocationAggregated teacher-specific matching lossesAggregated teacher-specific feature lossesGlobal/patch matching and teacher-dropping regularizationDynamic sample-wise teacher weightingTeacher-specific, position-level priorities derived from alignment difficulty, structural salience, and segmentation error
Student task organizationUnified image representation for downstream tasksDistilled visual representation for downstream robot policiesUnified image representation for semantic and geometric tasksRecognition training with downstream detection and segmentation transferShared 3D encoder jointly supporting lesion segmentation and multi-label presence prediction

The distinction is between controlling the importance of teacher losses and constructing local representation candidates whose compatibility is evaluated across teachers.

(2) What the Local Action Space Addresses

The action-space change addresses adaptive local composition in dense cross-dimensional transfer. Different regions of a medical volume can have different student deficiencies, structural characteristics, and lesion relevance. HiPPO makes both the local representation choice and the allocation of dense supervision explicit.

At position pp, the policy constructs a candidate:

f^p(g)=∑k=1Mak,p(g)ek,ap(g)∈{0,1}M.\hat f_p^{(g)}=\sum_{k=1}^{M}a_{k,p}^{(g)}e_k,\qquad a_p^{(g)}\in\{0,1\}^{M}.

The same candidate is evaluated in each teacher's native feature space:

qi,p(g)=cos⁡(hi(f^p(g)),ti,p),rp(g)=1K∑i=1Kqi,p(g).q_{i,p}^{(g)}=\cos\left(h_i(\hat f_p^{(g)}),t_{i,p}\right),\qquad r_p^{(g)}=\frac{1}{K}\sum_{i=1}^{K}q_{i,p}^{(g)}.

The student-conditioned policy learns local codebook compositions through group-relative comparisons of sampled rewards, rather than assigning teacher weights. Under nonlinear expert scoring, evaluating sampled compositions generally differs from evaluating their mean. The same feedback guides teacher-specific spatial priorities through a position-wise maximum of alignment difficulty, structural salience, and segmentation error, followed by unit-mean normalization that preserves coefficient budgets. HiPPO therefore coordinates what to compose and where to concentrate dense supervision. A shared 3D encoder jointly supports lesion segmentation and multi-label presence prediction, with organ-guided decoding when annotations are available. A separate training-only privileged branch transfers annotation-localized medical-VLM supervision through student soft tokens, independently of the visual rollout reward.

AuthorsOpenReview ↗

Relationship to Visual Multi-Teacher Distillation and the Contribution of HiPPO (2)

(3) Same-Protocol Comparisons with AM-RADIO-Style and Teacher-Weighted Distillation

We compare HiPPO with AM-RADIO-style equal-weight distillation and teacher-weighted distillation under the same experimental protocol. All values are percentages.

ModelmAUCMacro-BAccMacro-F1Macro-Dice
AM-RADIO-style, equal weight88.6282.0156.4149.92
Teacher-weighted distillation89.2882.6260.1051.74
HiPPO (ours)90.3383.5461.2952.41

HiPPO achieves the highest scores across all four metrics. Compared with AM-RADIO-style distillation, the gains are 1.71, 1.53, 4.88, and 2.49 percentage points in mAUC, Macro-BAcc, Macro-F1, and Macro-Dice, respectively. Against teacher-weighted distillation, the corresponding gains are 1.05, 0.92, 1.19, and 0.67 percentage points. These comparisons support the joint benefit of adaptive spatial supervision and learned prototype-subset selection beyond equal-weight and teacher-weighted distillation.

(4) Additional Experiments

We are currently running Theia and MTKD-RL comparisons that we had hoped to include in this rebuttal. Given the extensive additional experimental workload, we will report their results if they are completed before the rebuttal deadline. In our runs the MTKD-RL agent collapses to a uniform mixture (entropy ln⁡K\ln K, share ≈1/K\approx 1/K) after >2×105>2\times10^5 updates. Its released objective −r∑klog⁡wk-r\sum_k\log w_k over the simplex ∑kwk=1\sum_k w_k=1 is minimized uniquely at wk=1/Kw_k=1/K for any r>0r>0, and the scalar, teacher-agnostic reward cannot break this symmetry. The fused alignment term softmax(cos⁡)\mathrm{softmax}(\cos) is also near-uniform here, as the three experts have nearly equal student–teacher alignment (disagreement ≈0.05\approx 0.05). MTKD-RL thus reduces to equal-weight averaging, whereas HiPPO composes the codebook representation scored by the experts and learns non-uniform, per-position compositions.

AuthorsOpenReview ↗

The Necessity and Effectiveness of Reinforcement Learning

We thank the reviewer for highlighting this question. We use paired DeLong tests for per-organ AUC and paired case-level bootstraps for classification metrics and Dice. Here are our analysis.

(1) A Matched Control Identifies the Value of Sampled Optimization

Soft Mixture retains the codebook and gate parameterizations, teacher-specific projectors, three frozen experts, privileged-information term, and training protocol. It replaces sampled subsets with their expected composition, fˉ(x)=∑kσ(lk(x))ek\bar f(x)=\sum_k\sigma(l_k(x))e_k. Alignment then differentiates through the gates, without rollouts; the exploration entropy term is inactive in this deterministic variant. Against this matched control, HiPPO improves mAUC by +1.34+1.34 percentage points (pp), with a 95%95\% CI of [+0.48,+2.20][+0.48,+2.20] and p=0.0018p=0.0018. Because both variants retain the codebook and input-conditioned gates, this gain supports the sampled, group-relative optimization formulation beyond the shared representation architecture.

(2) Complementary Controls Support Learned, Input-Conditioned Selection

The controls examine different aspects of the formulation. All differences are expressed as HiPPO minus the control.

ControlModificationmAUC difference (pp)MacroF1 difference (pp)
Soft MixtureEvaluates the expected codebook composition+1.34+1.34; 95%95\% CI [+0.48,+2.20][+0.48,+2.20]; p=0.0018p=0.0018−0.30-0.30; 95%95\% CI [−2.43,+1.82][-2.43,+1.82]; p=0.77p=0.77
Deterministic projectionReplaces the codebook and gates with a continuous projection+0.64+0.64; 95%95\% CI [−0.46,+1.70][-0.46,+1.70]; p=0.243p=0.243+3.70+3.70; 95%95\% CI [+1.53,+5.86][+1.53,+5.86]; p=0.001p=0.001
Uniform random rolloutRetains sampling with gate probabilities fixed at 0.50.5+1.65+1.65; p=0.0004p=0.0004+3.14+3.14; p=0.0044p=0.0044
Input-independent dropoutRetains sampling without input-conditioned gating+1.08+1.08; p=0.020p=0.020+3.81+3.81; p=0.0006p=0.0006

HiPPO improves both mAUC and MacroF1 over random and input-independent sampling. These comparisons identify a contribution from learning which compositions to select, rather than stochastic feature variation alone. The deterministic-projection comparison provides complementary evidence, with its clearest gain in MacroF1 at the reported operating points.

(3) Benefits Extend to Lesion Segmentation

Relative to removing Codebook-GRPO, HiPPO gains +4.09+4.09 MacroF1 pp (p=0.0002p=0.0002) and +2.39+2.39 macro Dice points (95%95\% CI [+1.00,+3.84][+1.00,+3.84], p=0.0005p=0.0005), alongside a +0.84+0.84 mAUC change (p=0.063p=0.063). The liver Dice gain is +3.07+3.07 points (95%95\% CI [+0.84,+5.44][+0.84,+5.44], p=0.006p=0.006). Against Deterministic projection, HiPPO improves macro Dice by +1.75+1.75 points (95%95\% CI [+0.47,+3.07][+0.47,+3.07], p=0.0074p=0.0074). Together, the matched Soft Mixture comparison and the classification and segmentation results support the practical benefit of the sampled codebook-policy formulation. Its contribution is particularly evident in MacroF1 and voxel-overlap performance.

(4) Composition-Level Credit Explains the Optimization Rationale

At a fixed student state, hold the codebook and expert readouts fixed. Let βk=σ(lk(x))\beta_k=\sigma(l_k(x)), f^(a)=∑kakek\hat f(a)=\sum_k a_ke_k, and R(f)R(f) denote the expert-based compatibility reward. The expected-reward objective J=Ea[R(f^(a))]J=\mathbb E_a[R(\hat f(a))] has the exact gate derivative:

∂J∂lk=βk(1−βk) Ea−k[R(f−k+ek)−R(f−k)],\frac{\partial J}{\partial l_k}=\beta_k(1-\beta_k)\,\mathbb E_{a_{-k}}\left[R(f_{-k}+e_k)-R(f_{-k})\right],

where f−k=∑j≠kajejf_{-k}=\sum_{j\ne k}a_je_j.

An atom therefore receives credit according to its marginal reward contribution across the discrete contexts formed by the other selected atoms. Soft Mixture instead evaluates a directional derivative at one mean representation:

∂R(fˉ)∂lk=βk(1−βk) ∇R(fˉ)⊤ek.\frac{\partial R(\bar f)}{\partial l_k}=\beta_k(1-\beta_k)\,\nabla R(\bar f)^\top e_k.

Under nonlinear expert scoring, these signals are generally different. Codebook-GRPO constructs group-relative feedback from actual candidate evaluations, enabling selection learning to respond to combination-dependent usefulness.

This explains the rationale for sampled optimization, while the matched empirical comparisons establish its practical benefit.

(5) Multi-Seed Experiments

We report the mean results across training seeds. HiPPO achieves the highest mean on each reported metric. We acknowledge that, across the variants, some metrics were lower in the additional-seed runs than in the original experiments, reflecting variability across training runs.

ModelmAUCMacroF1Macro-BAccMacro-Dice
HiPPO89.2359.5582.6051.59
Deterministic projection88.6457.1681.8050.65
Soft Mixture88.5559.2181.7450.62
Reviewer_GzVHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HiPPO, an on-policy distillation framework that transfers knowledge from heterogeneous foundation models to a compact vision-only student for medical image understanding. HiPPO employs a student-conditioned policy to sample feature representations from a unified codebook and optimizes the policy using group-relative rewards from visual teachers. It further incorporates position-wise visual distillation and privileged semantic supervision from a medical vision-language model. The authors introduce MedBridge-Bench, covering over 18,000 volumetric CT, MRI, and PET/CT studies. Experiments demonstrate improved lesion classification and segmentation performance across multiple datasets and out-of-distribution settings, while ablation studies support the effectiveness of the proposed components.

Critical strengths

1.The paper presents a novel framework for distilling knowledge from heterogeneous foundation models into a compact vision-only student. By combining codebook-based policy optimization, position-wise visual distillation, and privileged semantic supervision, HiPPO enables adaptive knowledge transfer across teachers with different architectures, modalities, and representation spaces.

2.The paper provides extensive experimental validation across multiple datasets, imaging modalities, and clinical tasks. In particular, the out-of-distribution (OOD) evaluation on three independent external CT cohorts demonstrates strong generalization without additional fine-tuning or threshold recalibration. Comprehensive ablation studies further support the effectiveness of the proposed components.

Critical weaknesses

1.The necessity and effectiveness of reinforcement learning are not sufficiently justified. Although HiPPO employs codebook-based policy optimization, its improvements over simpler differentiable alternatives are relatively limited. In Table 5, removing Codebook-GRPO reduces AUC by only 0.84 percentage points, while Soft Mixture achieves a slightly higher F1 score than the full model. Furthermore, sensitivity analyses on key hyperparameters, such as the number of rollouts and codebook size, are missing. Additional experiments are needed to better justify the benefits of reinforcement learning over conventional differentiable approaches.

2.The paper lacks a thorough analysis of potential conflicts among heterogeneous teachers. HiPPO averages similarity-based rewards from different visual teachers without explicitly accounting for their varying reliability or disagreements. Although the ablation studies examine different teacher combinations, they do not investigate how conflicting or inaccurate teacher feedback affects policy optimization and downstream performance. Additional experiments on teacher disagreement and robustness would strengthen the proposed framework.

Questions for authors

Could the authors further clarify the necessity of reinforcement learning in HiPPO? Given the relatively limited improvements over simpler differentiable alternatives, what specific advantages does Codebook-GRPO offer that cannot be achieved through conventional differentiable feature fusion?

5 comments
AuthorsOpenReview ↗

The Necessity and Effectiveness of Reinforcement Learning

We thank the reviewer for highlighting this question. We use paired DeLong tests for per-organ AUC and paired case-level bootstraps for classification metrics and Dice. Here are our analysis.

(1) A Matched Control Identifies the Value of Sampled Optimization

Soft Mixture retains the codebook and gate parameterizations, teacher-specific projectors, three frozen experts, privileged-information term, and training protocol. It replaces sampled subsets with their expected composition, fˉ(x)=∑kσ(lk(x))ek\bar f(x)=\sum_k\sigma(l_k(x))e_k. Alignment then differentiates through the gates, without rollouts; the exploration entropy term is inactive in this deterministic variant. Against this matched control, HiPPO improves mAUC by +1.34+1.34 percentage points (pp), with a 95%95\% CI of [+0.48,+2.20][+0.48,+2.20] and p=0.0018p=0.0018. Because both variants retain the codebook and input-conditioned gates, this gain supports the sampled, group-relative optimization formulation beyond the shared representation architecture.

(2) Complementary Controls Support Learned, Input-Conditioned Selection

The controls examine different aspects of the formulation. All differences are expressed as HiPPO minus the control.

ControlModificationmAUC difference (pp)MacroF1 difference (pp)
Soft MixtureEvaluates the expected codebook composition+1.34+1.34; 95%95\% CI [+0.48,+2.20][+0.48,+2.20]; p=0.0018p=0.0018−0.30-0.30; 95%95\% CI [−2.43,+1.82][-2.43,+1.82]; p=0.77p=0.77
Deterministic projectionReplaces the codebook and gates with a continuous projection+0.64+0.64; 95%95\% CI [−0.46,+1.70][-0.46,+1.70]; p=0.243p=0.243+3.70+3.70; 95%95\% CI [+1.53,+5.86][+1.53,+5.86]; p=0.001p=0.001
Uniform random rolloutRetains sampling with gate probabilities fixed at 0.50.5+1.65+1.65; p=0.0004p=0.0004+3.14+3.14; p=0.0044p=0.0044
Input-independent dropoutRetains sampling without input-conditioned gating+1.08+1.08; p=0.020p=0.020+3.81+3.81; p=0.0006p=0.0006

HiPPO improves both mAUC and MacroF1 over random and input-independent sampling. These comparisons identify a contribution from learning which compositions to select, rather than stochastic feature variation alone. The deterministic-projection comparison provides complementary evidence, with its clearest gain in MacroF1 at the reported operating points.

(3) Benefits Extend to Lesion Segmentation

Relative to removing Codebook-GRPO, HiPPO gains +4.09+4.09 MacroF1 pp (p=0.0002p=0.0002) and +2.39+2.39 macro Dice points (95%95\% CI [+1.00,+3.84][+1.00,+3.84], p=0.0005p=0.0005), alongside a +0.84+0.84 mAUC change (p=0.063p=0.063). The liver Dice gain is +3.07+3.07 points (95%95\% CI [+0.84,+5.44][+0.84,+5.44], p=0.006p=0.006). Against Deterministic projection, HiPPO improves macro Dice by +1.75+1.75 points (95%95\% CI [+0.47,+3.07][+0.47,+3.07], p=0.0074p=0.0074). Together, the matched Soft Mixture comparison and the classification and segmentation results support the practical benefit of the sampled codebook-policy formulation. Its contribution is particularly evident in MacroF1 and voxel-overlap performance.

(4) Composition-Level Credit Explains the Optimization Rationale

At a fixed student state, hold the codebook and expert readouts fixed. Let βk=σ(lk(x))\beta_k=\sigma(l_k(x)), f^(a)=∑kakek\hat f(a)=\sum_k a_ke_k, and R(f)R(f) denote the expert-based compatibility reward. The expected-reward objective J=Ea[R(f^(a))]J=\mathbb E_a[R(\hat f(a))] has the exact gate derivative:

∂J∂lk=βk(1−βk) Ea−k[R(f−k+ek)−R(f−k)],\frac{\partial J}{\partial l_k}=\beta_k(1-\beta_k)\,\mathbb E_{a_{-k}}\left[R(f_{-k}+e_k)-R(f_{-k})\right],

where f−k=∑j≠kajejf_{-k}=\sum_{j\ne k}a_je_j.

An atom therefore receives credit according to its marginal reward contribution across the discrete contexts formed by the other selected atoms. Soft Mixture instead evaluates a directional derivative at one mean representation:

∂R(fˉ)∂lk=βk(1−βk) ∇R(fˉ)⊤ek.\frac{\partial R(\bar f)}{\partial l_k}=\beta_k(1-\beta_k)\,\nabla R(\bar f)^\top e_k.

Under nonlinear expert scoring, these signals are generally different. Codebook-GRPO constructs group-relative feedback from actual candidate evaluations, enabling selection learning to respond to combination-dependent usefulness.

This explains the rationale for sampled optimization, while the matched empirical comparisons establish its practical benefit.

(5) Multi-Seed Experiments

We report the mean results across training seeds. HiPPO achieves the highest mean on each reported metric. We acknowledge that, across the variants, some metrics were lower in the additional-seed runs than in the original experiments, reflecting variability across training runs.

ModelmAUCMacroF1Macro-BAccMacro-Dice
HiPPO89.2359.5582.6051.59
Deterministic projection88.6457.1681.8050.65
Soft Mixture88.5559.2181.7450.62
AuthorsOpenReview ↗

Hyperparameter Sensitivity

We report sensitivity analyses for rollout group size GG and codebook size MM, together with three additional factors: the number of codebooks RR, supervision density SS, and dense-distillation teacher-term aggregation.

(1) Rollout Group Size GG

GG is the number of sampled codebook subsets per position used to construct the group-relative advantage.

GGAUCBAccF1
2288.6281.3656.19
8890.3383.5461.29
161688.3281.1456.39

Among the tested values, G=8G=8 achieves the highest scores on all three metrics. Increasing the group size from 88 to 1616 does not improve performance in these runs. A larger group provides more candidate evaluations, but this does not guarantee better optimization under a finite training schedule. The table establishes the observed sensitivity to GG, rather than a mathematically necessary interior optimum.

(2) Codebook Size MM

MM is the number of atoms {ek}k=1M\{e_k\}_{k=1}^{M} available for composition.

MMAUCBAccF1
161688.1682.2061.32
323289.0682.4260.55
646490.3383.5461.29
12812888.6481.3157.64

M=64M=64 achieves the highest AUC and BAcc among the tested sizes, whereas M=16M=16 has a slightly higher F1 point estimate, by 0.03 percentage points. Increasing MM to 128128 lowers all three metrics in the reported run. Thus, a larger set of available compositions does not automatically translate into better learned representations.

(3) Number of Codebooks RR

We vary RR while keeping M=64M=64 and K=3K=3 fixed. The codebooks are read by separate gate heads and share the same expert-based reward definition.

RRAUCBAccF1
1190.3383.5461.29
2289.3482.9161.92
3383.0475.1751.01

Two codebooks improve the F1 point estimate by 0.63 percentage points, but reduce AUC and BAcc. Three codebooks produce substantially lower scores in this run. The results therefore do not show a consistent benefit from adding codebooks, and we retain the submitted single-codebook configuration.

(4) Dense Supervision Slices SS

SS is the number of axial positions per volume receiving the position-wise distillation term.

SSAUCBAccF1
8889.3282.6160.05
161689.4282.9061.16
323290.3383.5461.29
646486.3878.9254.75

The submitted setting, S=32S=32, achieves the highest scores among the tested densities. Increasing supervision density to 6464 lowers all three metrics in the reported run. These results show that supervising more positions is not automatically beneficial; they do not establish whether the decline arises from optimization scaling, redundancy between positions, or another mechanism.

(5) Teacher-Term Aggregation in Dense Distillation

The default dense-distillation objective averages teacher-specific terms:

Ldense=1K∑i=1K1P∑pωi,pδi,p.\mathcal L_{\mathrm{dense}}=\frac{1}{K}\sum_{i=1}^{K}\frac{1}{P}\sum_p\omega_{i,p}\delta_{i,p}.

We compare equal aggregation with a compatibility-conditioned selector:

ρi,p(τ)=exp⁡(ci,p/τ)∑j=1Kexp⁡(cj,p/τ),\rho_{i,p}(\tau)=\frac{\exp(c_{i,p}/\tau)}{\sum_{j=1}^{K}\exp(c_{j,p}/\tau)},

where ci,pc_{i,p} is the current student–teacher compatibility at position pp. The selector is computed under stop-gradient:

Ldense,τ=1P∑p∑i=1Ksg⁡[ρi,p(τ)] ωi,pδi,p.\mathcal L_{\mathrm{dense},\tau}=\frac{1}{P}\sum_p\sum_{i=1}^{K}\operatorname{sg}[\rho_{i,p}(\tau)]\,\omega_{i,p}\delta_{i,p}.

These variants retain the teacher-specific spatial priorities and leave the scalar policy reward unchanged: rp(g)=1K∑iqi,p(g)r_p^{(g)}=\frac{1}{K}\sum_iq_{i,p}^{(g)}.

Alignment-loss aggregationmAUCMacro-BAccMacroF1
Equal mean90.3383.5461.29
Softmax, τ=0.1\tau=0.188.7781.7257.83
Softmax, τ=0.5\tau=0.585.1477.8054.04
Softmax, τ=1.0\tau=1.088.6081.9257.19
Softmax, τ=5.0\tau=5.088.9081.9158.24

Equal aggregation achieves the highest reported scores among the tested variants, with mAUC advantages of 1.42–5.19 percentage points. These results favour equal teacher-term aggregation combined with teacher-specific spatial priorities.

AuthorsOpenReview ↗

Teacher Disagreement, Aggregation, and Downstream Performance (1)

We thank the reviewer for highlighting this question. We analyze teacher-specific dense-distillation gradients, their association with downstream predictions, and controls on teacher-term aggregation and spatial supervision. These measurements characterize the evaluated teacher pool, while a separate policy-credit decomposition explains how teacher evaluations enter selection learning.

(1) Teacher-Specific Supervision and Policy Rewards

Each visual teacher retains its own readout head and detached target. The deterministic dense-distillation objective combines separate teacher-specific discrepancies:

Ldense=1K∑i=1K1P∑pωi,pδi,pdense,1P∑pωi,p=1.\mathcal L_{\mathrm{dense}}=\frac{1}{K}\sum_{i=1}^{K}\frac{1}{P}\sum_p\omega_{i,p}\delta_{i,p}^{\mathrm{dense}},\qquad \frac{1}{P}\sum_p\omega_{i,p}=1.

Here, δi,pdense\delta_{i,p}^{\mathrm{dense}} measures the discrepancy of the deterministic student feature against teacher ii, and ωi,p\omega_{i,p} allocates its supervision across positions. This term is distinct from the alignment of hard-sampled codebook compositions.

The policy reward separately combines evaluations of each sampled composition:

rp(g)=1K∑i=1Kqi,p(g).r_p^{(g)}=\frac{1}{K}\sum_{i=1}^{K}q_{i,p}^{(g)}.

On the full 926-case evaluation cohort, the proposed three-teacher model improves over its best independently trained two-teacher subset by +1.11+1.11 mAUC pp, with a paired-bootstrap 95%95\% CI of [+0.27,+1.96][+0.27,+1.96] and p=0.010p=0.010. This supports the joint utility of the selected teacher constraints.

For independent Bernoulli gates, the logit-space score is ∇lplog⁡πϕ(a(g)∣xp)=a(g)−βp\nabla_{\mathbf l_p}\log\pi_\phi(a^{(g)}\mid x_p)=a^{(g)}-\boldsymbol\beta_p, where βk,p=σ(lk(xp))\beta_{k,p}=\sigma(l_k(x_p)). Using a common aggregate-reward normalization, each teacher contributes:

vi,p=1G∑gqi,p(g)−qˉi,pσr,p+ϵ(a(g)−βp),\mathbf v_{i,p}=\frac{1}{G}\sum_g\frac{q_{i,p}^{(g)}-\bar q_{i,p}}{\sigma_{r,p}+\epsilon}\left(a^{(g)}-\boldsymbol\beta_p\right),

where qˉi,p\bar q_{i,p} is the teacher's group-mean score and σr,p\sigma_{r,p} is the standard deviation of the aggregate rewards. The score-function credit satisfies up=1K∑ivi,p\mathbf u_p=\frac{1}{K}\sum_i\mathbf v_{i,p}. This finite-sample identity locates teacher contributions in gate-selection credit; the following measurements examine the raw dense-distillation constraints.

(2) Pairwise Conflict and Aggregate Opposition

We measure gradient geometry on the same 370 lesion-positive evaluation cases for each teacher pool:

gi=∇B[1P∑pδi,pdense].\mathbf g_i=\nabla_{\mathbf B}\left[\frac{1}{P}\sum_p\delta_{i,p}^{\mathrm{dense}}\right].

All diagnostic vectors occupy the same student-bottleneck space, before spatial-priority weighting. For nonzero gradients, we measure pairwise angles and cancellation:

γij=arccos⁡(⟨gi,gj⟩∥gi∥ ∥gj∥),C=1−∥∑igi∥∑i∥gi∥.\gamma_{ij}=\arccos\left(\frac{\langle\mathbf g_i,\mathbf g_j\rangle}{\|\mathbf g_i\|\,\|\mathbf g_j\|}\right),\qquad C=1-\frac{\left\|\sum_i\mathbf g_i\right\|}{\sum_i\|\mathbf g_i\|}.

Pairwise conflict is defined by γij>90∘\gamma_{ij}>90^\circ. For the aggregate raw-gradient direction G=1K∑igi\mathbf G=\frac{1}{K}\sum_i\mathbf g_i, teacher ii is opposed exactly when:

Ti=1+∑j≠i∥gj∥∥gi∥cos⁡γij=K⟨G,gi⟩∥gi∥2<0.T_i=1+\sum_{j\ne i}\frac{\|\mathbf g_j\|}{\|\mathbf g_i\|}\cos\gamma_{ij}=\frac{K\langle\mathbf G,\mathbf g_i\rangle}{\|\mathbf g_i\|^2}<0.

Thus, aggregate opposition depends jointly on directional disagreement and relative contribution magnitude.

Teacher poolMean anglePair–case conflict (%)Mean cancellation [95% CI]Maximum cancellationNorm ratio: median / 95th percentileOutvoted cases (%)mAUC (%)
Proposed: DINOv3, BiomedCLIP, MedSAM287.4∘87.4^\circ11.80.378 [0.376, 0.380]0.4361.81 / 2.730.090.33
CLIP-family: BiomedCLIP, UniMedCLIP, SigLIP83.8∘83.8^\circ4.40.359 [0.357, 0.361]0.4001.77 / 2.360.088.51
SSL-family: DINOv3, I-JEPA, LVM-Med87.2∘87.2^\circ18.10.315 [0.312, 0.319]0.4786.71 / 11.421.189.28

The proposed pool combines directional diversity with comparatively balanced magnitudes. Across the 370 cases, its aggregate raw-gradient direction has a non-negative projection onto every measured teacher gradient. The SSL-family pool exhibits greater magnitude imbalance and some aggregate opposition.

For reference, three equal-magnitude mutually orthogonal gradients give Corth=1−1/3≈0.4226C_{\mathrm{orth}}=1-1/\sqrt{3}\approx0.4226. We interpret cancellation alongside angles and norm ratios.

Comments on the paper

12 comments
AuthorsOpenReview ↗

The Necessity and Effectiveness of Reinforcement Learning

We thank the reviewer for highlighting this question. We use paired DeLong tests for per-organ AUC and paired case-level bootstraps for classification metrics and Dice. Here are our analysis.

(1) A Matched Control Identifies the Value of Sampled Optimization

Soft Mixture retains the codebook and gate parameterizations, teacher-specific projectors, three frozen experts, privileged-information term, and training protocol. It replaces sampled subsets with their expected composition, fˉ(x)=∑kσ(lk(x))ek\bar f(x)=\sum_k\sigma(l_k(x))e_k. Alignment then differentiates through the gates, without rollouts; the exploration entropy term is inactive in this deterministic variant. Against this matched control, HiPPO improves mAUC by +1.34+1.34 percentage points (pp), with a 95%95\% CI of [+0.48,+2.20][+0.48,+2.20] and p=0.0018p=0.0018. Because both variants retain the codebook and input-conditioned gates, this gain supports the sampled, group-relative optimization formulation beyond the shared representation architecture.

(2) Complementary Controls Support Learned, Input-Conditioned Selection

The controls examine different aspects of the formulation. All differences are expressed as HiPPO minus the control.

ControlModificationmAUC difference (pp)MacroF1 difference (pp)
Soft MixtureEvaluates the expected codebook composition+1.34+1.34; 95%95\% CI [+0.48,+2.20][+0.48,+2.20]; p=0.0018p=0.0018−0.30-0.30; 95%95\% CI [−2.43,+1.82][-2.43,+1.82]; p=0.77p=0.77
Deterministic projectionReplaces the codebook and gates with a continuous projection+0.64+0.64; 95%95\% CI [−0.46,+1.70][-0.46,+1.70]; p=0.243p=0.243+3.70+3.70; 95%95\% CI [+1.53,+5.86][+1.53,+5.86]; p=0.001p=0.001
Uniform random rolloutRetains sampling with gate probabilities fixed at 0.50.5+1.65+1.65; p=0.0004p=0.0004+3.14+3.14; p=0.0044p=0.0044
Input-independent dropoutRetains sampling without input-conditioned gating+1.08+1.08; p=0.020p=0.020+3.81+3.81; p=0.0006p=0.0006

HiPPO improves both mAUC and MacroF1 over random and input-independent sampling. These comparisons identify a contribution from learning which compositions to select, rather than stochastic feature variation alone. The deterministic-projection comparison provides complementary evidence, with its clearest gain in MacroF1 at the reported operating points.

(3) Benefits Extend to Lesion Segmentation

Relative to removing Codebook-GRPO, HiPPO gains +4.09+4.09 MacroF1 pp (p=0.0002p=0.0002) and +2.39+2.39 macro Dice points (95%95\% CI [+1.00,+3.84][+1.00,+3.84], p=0.0005p=0.0005), alongside a +0.84+0.84 mAUC change (p=0.063p=0.063). The liver Dice gain is +3.07+3.07 points (95%95\% CI [+0.84,+5.44][+0.84,+5.44], p=0.006p=0.006). Against Deterministic projection, HiPPO improves macro Dice by +1.75+1.75 points (95%95\% CI [+0.47,+3.07][+0.47,+3.07], p=0.0074p=0.0074). Together, the matched Soft Mixture comparison and the classification and segmentation results support the practical benefit of the sampled codebook-policy formulation. Its contribution is particularly evident in MacroF1 and voxel-overlap performance.

(4) Composition-Level Credit Explains the Optimization Rationale

At a fixed student state, hold the codebook and expert readouts fixed. Let βk=σ(lk(x))\beta_k=\sigma(l_k(x)), f^(a)=∑kakek\hat f(a)=\sum_k a_ke_k, and R(f)R(f) denote the expert-based compatibility reward. The expected-reward objective J=Ea[R(f^(a))]J=\mathbb E_a[R(\hat f(a))] has the exact gate derivative:

∂J∂lk=βk(1−βk) Ea−k[R(f−k+ek)−R(f−k)],\frac{\partial J}{\partial l_k}=\beta_k(1-\beta_k)\,\mathbb E_{a_{-k}}\left[R(f_{-k}+e_k)-R(f_{-k})\right],

where f−k=∑j≠kajejf_{-k}=\sum_{j\ne k}a_je_j.

An atom therefore receives credit according to its marginal reward contribution across the discrete contexts formed by the other selected atoms. Soft Mixture instead evaluates a directional derivative at one mean representation:

∂R(fˉ)∂lk=βk(1−βk) ∇R(fˉ)⊤ek.\frac{\partial R(\bar f)}{\partial l_k}=\beta_k(1-\beta_k)\,\nabla R(\bar f)^\top e_k.

Under nonlinear expert scoring, these signals are generally different. Codebook-GRPO constructs group-relative feedback from actual candidate evaluations, enabling selection learning to respond to combination-dependent usefulness.

This explains the rationale for sampled optimization, while the matched empirical comparisons establish its practical benefit.

(5) Multi-Seed Experiments

We report the mean results across training seeds. HiPPO achieves the highest mean on each reported metric. We acknowledge that, across the variants, some metrics were lower in the additional-seed runs than in the original experiments, reflecting variability across training runs.

ModelmAUCMacroF1Macro-BAccMacro-Dice
HiPPO89.2359.5582.6051.59
Deterministic projection88.6457.1681.8050.65
Soft Mixture88.5559.2181.7450.62
AuthorsOpenReview ↗

Teacher Disagreement, Aggregation, and Downstream Performance (1)

We thank the reviewer for highlighting this question. We analyze teacher-specific dense-distillation gradients, their association with downstream predictions, and controls on teacher-term aggregation and spatial supervision. These measurements characterize the evaluated teacher pool, while a separate policy-credit decomposition explains how teacher evaluations enter selection learning.

(1) Teacher-Specific Supervision and Policy Rewards

Each visual teacher retains its own readout head and detached target. The deterministic dense-distillation objective combines separate teacher-specific discrepancies:

Ldense=1K∑i=1K1P∑pωi,pδi,pdense,1P∑pωi,p=1.\mathcal L_{\mathrm{dense}}=\frac{1}{K}\sum_{i=1}^{K}\frac{1}{P}\sum_p\omega_{i,p}\delta_{i,p}^{\mathrm{dense}},\qquad \frac{1}{P}\sum_p\omega_{i,p}=1.

Here, δi,pdense\delta_{i,p}^{\mathrm{dense}} measures the discrepancy of the deterministic student feature against teacher ii, and ωi,p\omega_{i,p} allocates its supervision across positions. This term is distinct from the alignment of hard-sampled codebook compositions.

The policy reward separately combines evaluations of each sampled composition:

rp(g)=1K∑i=1Kqi,p(g).r_p^{(g)}=\frac{1}{K}\sum_{i=1}^{K}q_{i,p}^{(g)}.

On the full 926-case evaluation cohort, the proposed three-teacher model improves over its best independently trained two-teacher subset by +1.11+1.11 mAUC pp, with a paired-bootstrap 95%95\% CI of [+0.27,+1.96][+0.27,+1.96] and p=0.010p=0.010. This supports the joint utility of the selected teacher constraints.

For independent Bernoulli gates, the logit-space score is ∇lplog⁡πϕ(a(g)∣xp)=a(g)−βp\nabla_{\mathbf l_p}\log\pi_\phi(a^{(g)}\mid x_p)=a^{(g)}-\boldsymbol\beta_p, where βk,p=σ(lk(xp))\beta_{k,p}=\sigma(l_k(x_p)). Using a common aggregate-reward normalization, each teacher contributes:

vi,p=1G∑gqi,p(g)−qˉi,pσr,p+ϵ(a(g)−βp),\mathbf v_{i,p}=\frac{1}{G}\sum_g\frac{q_{i,p}^{(g)}-\bar q_{i,p}}{\sigma_{r,p}+\epsilon}\left(a^{(g)}-\boldsymbol\beta_p\right),

where qˉi,p\bar q_{i,p} is the teacher's group-mean score and σr,p\sigma_{r,p} is the standard deviation of the aggregate rewards. The score-function credit satisfies up=1K∑ivi,p\mathbf u_p=\frac{1}{K}\sum_i\mathbf v_{i,p}. This finite-sample identity locates teacher contributions in gate-selection credit; the following measurements examine the raw dense-distillation constraints.

(2) Pairwise Conflict and Aggregate Opposition

We measure gradient geometry on the same 370 lesion-positive evaluation cases for each teacher pool:

gi=∇B[1P∑pδi,pdense].\mathbf g_i=\nabla_{\mathbf B}\left[\frac{1}{P}\sum_p\delta_{i,p}^{\mathrm{dense}}\right].

All diagnostic vectors occupy the same student-bottleneck space, before spatial-priority weighting. For nonzero gradients, we measure pairwise angles and cancellation:

γij=arccos⁡(⟨gi,gj⟩∥gi∥ ∥gj∥),C=1−∥∑igi∥∑i∥gi∥.\gamma_{ij}=\arccos\left(\frac{\langle\mathbf g_i,\mathbf g_j\rangle}{\|\mathbf g_i\|\,\|\mathbf g_j\|}\right),\qquad C=1-\frac{\left\|\sum_i\mathbf g_i\right\|}{\sum_i\|\mathbf g_i\|}.

Pairwise conflict is defined by γij>90∘\gamma_{ij}>90^\circ. For the aggregate raw-gradient direction G=1K∑igi\mathbf G=\frac{1}{K}\sum_i\mathbf g_i, teacher ii is opposed exactly when:

Ti=1+∑j≠i∥gj∥∥gi∥cos⁡γij=K⟨G,gi⟩∥gi∥2<0.T_i=1+\sum_{j\ne i}\frac{\|\mathbf g_j\|}{\|\mathbf g_i\|}\cos\gamma_{ij}=\frac{K\langle\mathbf G,\mathbf g_i\rangle}{\|\mathbf g_i\|^2}<0.

Thus, aggregate opposition depends jointly on directional disagreement and relative contribution magnitude.

Teacher poolMean anglePair–case conflict (%)Mean cancellation [95% CI]Maximum cancellationNorm ratio: median / 95th percentileOutvoted cases (%)mAUC (%)
Proposed: DINOv3, BiomedCLIP, MedSAM287.4∘87.4^\circ11.80.378 [0.376, 0.380]0.4361.81 / 2.730.090.33
CLIP-family: BiomedCLIP, UniMedCLIP, SigLIP83.8∘83.8^\circ4.40.359 [0.357, 0.361]0.4001.77 / 2.360.088.51
SSL-family: DINOv3, I-JEPA, LVM-Med87.2∘87.2^\circ18.10.315 [0.312, 0.319]0.4786.71 / 11.421.189.28

The proposed pool combines directional diversity with comparatively balanced magnitudes. Across the 370 cases, its aggregate raw-gradient direction has a non-negative projection onto every measured teacher gradient. The SSL-family pool exhibits greater magnitude imbalance and some aggregate opposition.

For reference, three equal-magnitude mutually orthogonal gradients give Corth=1−1/3≈0.4226C_{\mathrm{orth}}=1-1/\sqrt{3}\approx0.4226. We interpret cancellation alongside angles and norm ratios.

AuthorsOpenReview ↗

Teacher Disagreement, Aggregation, and Downstream Performance (2)

(3) Directional Structure Relative to Sign Randomization

We compare the observed directions with a sign-flip reference that preserves coordinate magnitudes and dimensionality. Symmetric coordinate sign flips give an expected angle of 90∘90^\circ. A separate 64-case organ-balanced diagnostic cohort at the final checkpoint gives the following results, using 200 sign-flip draws and 2,000 bootstrap draws:

Teacher poolMean angle95% CIObservations outside their own null band (%)
Proposed86.94∘86.94^\circ[86.49∘,87.34∘][86.49^\circ,87.34^\circ]87.0
CLIP-family83.73∘83.73^\circ[82.81∘,84.65∘][82.81^\circ,84.65^\circ]90.6
SSL-family86.82∘86.82^\circ[86.38∘,87.24∘][86.38^\circ,87.24^\circ]92.2

The gradients contain directional structure relative to magnitude-preserving sign randomization, with positive average agreement alongside substantial directional diversity.

We also flag near-null observations with strongly unequal magnitudes, using κij=min⁡(∥gi∥,∥gj∥)/max⁡(∥gi∥,∥gj∥)<0.1\kappa_{ij}=\min(\|\mathbf g_i\|,\|\mathbf g_j\|)/\max(\|\mathbf g_i\|,\|\mathbf g_j\|)<0.1. Each cohort is assessed against its own null bands.

On the 370-case evaluation cohort, the proposed and CLIP-family pairs have zero joint-flag rates. Two SSL-family pairs, DINOv3–I-JEPA and DINOv3–LVM-Med, each have a 1.1% rate. The appendix will report the full pair-specific results and diagnostic definitions.

(4) Association with Downstream Predictions

The pooled worst-pair gradient cosine has a modest positive association with the prediction margin m=p^organ−θorganm=\hat p_{\mathrm{organ}}-\theta_{\mathrm{organ}} (Spearman ρ=+0.134\rho=+0.134, p=0.010p=0.010). Conflict frequency differs across organs, so we additionally report stratified associations. Within each organ, subtracting its fixed threshold preserves rank correlations; these values therefore apply to both the prediction margin and the own-organ prediction score.

OrganSpearman correlation with worst-pair gradient cosineNominal ppHolm-adjusted pp
Kidney+0.194+0.1940.0100.030
Pancreas+0.226+0.2260.0290.059
Liver+0.096+0.0960.3430.343

The associations are positive and organ-dependent, with the kidney result retaining statistical support after Holm adjustment. Across the tested pools, the lowest conflict rate does not coincide with the highest mAUC. Directional geometry, contribution magnitude, and downstream utility therefore provide complementary assessments of the teacher pool.

(5) Controls on Teacher-Term Aggregation

We also compare equal dense-distillation aggregation with a compatibility-conditioned selector. Equal aggregation achieves higher mAUC than all four selectors, by 1.42–5.19 pp, and higher organ-specific AUC in 11 of 12 comparisons. With unit-mean spatial priorities, equal aggregation preserves each teacher's average coefficient budget. Routing additionally favors currently better-matched teachers, whereas the discrepancy-based spatial priority emphasizes positions requiring further alignment within each teacher's supervision. The uniform-temperature limit recovers equal aggregation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.