acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Curriculum Learning for Multimodal Diffusion Language Models in Biology

Abstract

Biological modeling is inherently multimodal, spanning perturbations, cellular phenotypes, biomedical images, spatial structures, and textual descriptions. Multimodal diffusion language models provide a flexible, noncausal framework in these settings. However, multimodal diffusion introduces a training challenge that is largely absent from the usual unimodal view of diffusion: different prediction directions and modality-specific corruption states define distinct training problems. We introduce MaDAC, an adaptive curriculum learning method that allocates training across these multimodal prediction problems. MaDAC uses an outer curriculum to adapt prediction routes and modality-specific corruption states to the learner state, and an inner curriculum to reweight targets, while a reference distribution preserves coverage. Our analysis shows that standard corruption schedules can severely under-sample training states with asymmetric corruption across modalities, and that adaptive allocation can help when their relative utility changes during learning. We evaluate MaDAC on multimodal biological tasks spanning perturbation-phenotype modeling, biomedical imaging, and spatial segmentation. Across settings, MaDAC significantly improves over standard corruption and fixed-allocation curricula, demonstrating the value of adaptive curriculum allocation for multimodal diffusion language models for biological tasks.

Primary area:
generative models
Keywords:
Biology, Multimodal models, Diffusion Language Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_yNzhOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

In multi-modal diffusion models, training needs to choose which tokens of which modality are masked and which are visible. This process implicitly allocates the training budget across different modalities, and usually this is uniform at random. The paper proposes a technique, Multimodal DLM Adaptive Curriculum learning (MaDAC), that explicitly allocates training across the different modalities. The algorithm has an outer and an inner step. The outer step chooses which direction to predict (e.g., from molecule modality to cell image modality) and how much of each modality to mask. It favors choices on which the model's loss on fixed probe data is dropping fastest. A fixed fraction of directions is always kept to prevent forgetting. The inner step re-weights the hidden tokens inside one example, by default giving more weight to tokens the model is already confident about. MaDAC is evaluated on three biology tasks: linking compounds to cell morphology (JUMP), medical image segmentation from a text prompt (BioMed), and tissue segmentation on organ slides (HuBMAP). The authors test their algorithm on two different models: a pretrained LLaDA2.0-Uni (16B) fine-tuned with LoRA, and a 147M model (UniDisc) trained from scratch. MaDAC beats the same model trained with standard random masking in most settings. For example, with LLaDA2, BioMed Dice goes from 0.469 to 0.616 on held-out sources and from 0.775 to 0.860 in-distribution, and HuBMAP held-out Dice goes from 0.387 to 0.431; with UniDisc, JUMP SMILES→morphology Hits@10 goes from 0.486 to 0.564. Gains are smaller elsewhere, and LLaDA2-based JUMP results are mixed.

Critical strengths

I found the idea of treating the multi-modal diffusion training as an allocation problem and building an adaptive curriculum around it interesting and, as far as I know, new.

Some results are strong, especially on BioMed source-held-out segmentation with the pre-trained Llada backbone.

The authors shared their code which was helpful to better understand some aspects of the experimental settings.

Critical weaknesses

I find the overall idea interesting, but a few concerns related to the experiments prevent me from understanding if the algorithm is actually useful in practice. Specifically: the experiments do not seem to clearly test the adaptive part of the algorithm described (and the shared code suggests the reported runs may not use it), and I have concerns related to the fairness of the comparison with baselines. Additionally, from a reproducibly point of view, some information is currently missing.

The paper claims that the adaptive allocation causes significant gains across settings. I find this claim only partially supported by the results shown.

  • Without a proper ablation I cannot tell if the gains come from hiding each modality at a separate rate, or from adapting those rates during training (given that MaDAC does both and no baseline or ablation does only one of them). The missing baseline that would answer this is simply to draw each modality's rate from a fixed, separate distribution, with no adaptation. This would be especially relevant on BioMed, where the gains are largest, but it has no ablations at all. Even the ablations on JUMP do not test this, as the only non-adaptive version (“fixed-allocation”) is the uniform baseline itself (identical numbers), which uses one constant setting, equal for both modalities at every step.
  • On the HuBMAP task, it seems to me that MaDAC does not use its ability to choose the path, since the task is fixed (?) to predict the mask from the image and the organ name. If that is the case, then the outer step boils down to simply showing harder-samples more often, which is a known technique. If this is correct, then HuBMAP cannot be used to support the paper’s main claim about allocating corruption across modalities.
  • In general, results are mix rather than showing consistent and significant gain across settings. I would say that they gains are large and significant for BioMed with Lllada2, small or within standard deviation in many of the other settings, and slightly negative for Llada2 on 2 of 3 JUMP retrieval tasks. So "significant across settings" is only partially supported.
  • The code shared raises more concerns about the paper since what seems implemented there is not what is described in the paper. Specifically none of the three tasks seems to support the claim that “MaDAC adapts training distribution in response to the learner state”:
    • BioMed's outer level seems to use a pre-set anneal not the algorithm presented in the paper.
    • JUMP's seems to use a fixed 50/50 route mix, not the dynamic algorithm presented in the paper.
    • HuBMAP has a single route (this was clear also from the paper and confirmed in the code).

I am not sure the set up is completely fair to the baselines

  • On HuBMAP, it seems that only MaDAC gets a warm-up phase that is used to train the decoder first. So it is unclear how much of the gain comes from this additional training budget. This additional budget should be given to the baselines too as additional training steps (or as warm-up exactly as for MaDAC).
  • MaDAC's controller measures progress on "fixed probe data" but the paper only says it uses “training or calibration data” and not where they are coming from. If it is the calibration/validation set, MaDAC receives a training-time signal from data the baselines don’t receive.

Reproducibility concerns

  • Hyperparameter values or tuning budgets are not reported for MaDAC nor for uniform / easy-to-hard. This prevents reproduction from the paper alone and makes it impossible to rule out that gains come from extra tuning.
  • Tables report mean and standard deviation, but the paper never says how many seeds were run or what the standard deviation is computed over.

Not reasons for rejection, but worth a discussion.

  • Every time I see a curriculum-based work, I am reminded of the paper "When Do Curricula Work?" (2021), which found that curricula help mainly with limited training budgets or noisy data, but almost all benefit disappear with enough compute. Is it the case here too? I guess in order to test this the authors would need: - UniDisc at several sizes trained from scratch, with the same budget; - LLaDA2 with full fine-tuning and longer budgets, to see whether the MaDAC–uniform gap closes; - gap vs. training steps, i.e., Fig. 9-style curves for BioMed and JUMP.
  • The format of the paper is very dense: figures and table are packed closely, fonts on figures are too small. All this makes reading the paper a difficult experience.
Questions for authors
  • Could you run the ablation described above (fixed, separate per-modality rates with no adaptation) on BioMed and JUMP)?
  • Could you explain the discrepancy between the algorithm explained in the paper and the config files that describe the algorithm for the three tasks?
  • Can you report results on HuBMAP for a uniform baseline with the same decoder warm-up as MaDAC?
  • Can you clarify where the “fixed probe data" used by MaDAC comes from, and confirm that the setting is fair to the baselines?
  • Can you list all hyperparameter values and how you found them, for MaDAC and for the baselines?
  • Could you clarify how many seeds were run and what is the standard deviation is computed over?
  • Can you comment on my concern related to the benefit of curricula disappearing as scale (of data or compute) increases?
13 comments
AuthorsOpenReview ↗

Response to Reviewer yNzh

Dear Reviewer yNzh,

Thank you for your constructive suggestions and insightful comments! We thank the reviewer for the questions about implementation, attribution, fairness and scaling. G.2 defines the common controls; G.5–G.7 report the task results; G.8 explains their implications and data roles. The responses below connect each question to those comparisons. We also provide further clarifications and refinements in the general response and paper revision.

Q1: Fixed, separate per-modality rates without adaptation, on BioMed and JUMP

Could you run the ablation described above... on BioMed and JUMP?

A2 supplies the requested fixed, separate per-modality rates without adaptation. On BioMed, B1 − A2 improves held-out global Dice by +0.046 / +0.085 at 1× and +0.084 / +0.018 at 2×, in source/target-calibration order. The separately trained IID adapter gains +0.044 [95% CI +0.042, +0.046] (G.7, Tables R1.1–R1.5 and R1-C).

The A1–A4 factorial design answers whether separate rates or their adaptation helps. Adapting separate rates (A4 − A2) gives +0.112 / +0.096 at 1×, +0.078 / +0.048 at 2×, and +0.027 on the separately trained IID adapter. With both policies adapted, separate rates (A4 − A3) give +0.051 / +0.065 on held-out data. Separate fixed rates (A2 − A1) show no consistent held-out gain. On BioMed, separate rates also help (Table R1-R).

JUMP's matched A2 control and full component comparisons appear in G.5 (R2-U/R2-L). B1 − A2 demonstrates all three retrieval endpoints as positive on both backbones. Rate-only adaptation is also compared, showing that its contribution depends on the task and backbone.

Q2: How the submitted configurations relate to the algorithm in the paper

BioMed's outer level seems to use a pre-set anneal...

We ran Algorithm 1 on BioMed as B1. It learns route allocation and bounded rate policies using absolute progress and the same-batch reward, with reference mass 0.20 and Eq. 6 target weighting. B1 improves over fixed independent A2 under both calibration rules, at both training budgets, and on the separately trained IID adapter (G.7).

The revised JUMP comparison uses the same B1 definition. HuBMAP-MR supplies four prediction routes under a common read-out/loss interface and additionally uses example sampling.

Q3 and the hard-example concern: matched decoder warm-up, and what HuBMAP measures

Can you report results on HuBMAP for a uniform baseline with the same decoder warm-up as MaDAC? … the outer step boils down to simply showing harder samples more often.

HuBMAP-MR gives all matched arms the same training schedule and learning-rate warm-up, with the adapter/backbone trained from the first update. The common interface and schedule are introduced in G.2, and G.6 contains the complete R3 results. The revised comparison therefore matches the training schedule across curriculum arms.

To address the hard-example explanation, B1–B4 retains example sampling and target weighting while fixing allocation: B1 has higher mean image Dice on both backbones. B5/B6 separately remove target weighting/example sampling, and allocation-only B7 falls below A2 in mean image Dice. Thus the evidence identifies what allocation adds to the combined policy and how the gain depends on its supporting factors. G.6 reports these ablations alongside segmentation-only R2/R3.

Mixed results and “significant across settings”

results are mixed rather than showing consistent and significant gain across settings.

The revised claim specifies the endpoint and comparator. B1 improves BioMed held-out global Dice against A2 under both calibration rules and training budgets, and improves separately trained IID Dice. It also achieves higher mean source-held-out image Dice on HuBMAP-MR and the highest means on all five JUMP endpoints with both backbones. JUMP's reported comparisons resolve all three retrieval endpoints.

The component results explain where the gains arise. BioMed's A1–A4 design identifies rate adaptation as the key contributor, and B1 − B4 improves AUROC under both calibration rules and separately trained IID Dice. HuBMAP-MR's B1 − B4 comparison shows higher mean image Dice with example sampling and target weighting fixed. G.7–G.8 give the endpoint-specific interpretation.

Q4: Where the probe data come from, and fairness to baselines

MaDAC's controller measures progress on “fixed probe data” …

The training protocol prescribes fixed training-split probes., with exact prescribed settings in R6-A–R6-D below. The comparison uses the stated validation/calibration selection recipe for each arm. Training probes, model selection and the utility audit thus have explicit, separate data roles.

AuthorsOpenReview ↗

Response to Reviewer yNzh

Q5: All hyperparameters and how they were found

Tables R6-A–R6-F list the training recipes, MaDAC parameters, baseline switches and selection procedures for JUMP, HuBMAP-MR and BioMed. Shared optimization settings apply to all arms within each task/backbone; Table R6-E specifies the curriculum changes. The implementation records describe the shared training recipes and the common controller constants as fixed in advance. Each table identifies the source of its settings.

Table R6-A — shared MaDAC controller and target-weight parameters

ParameterSetting
Reference mass ρ0.20; q = 0.8π + 0.2p₀. Stationary arms draw every action from p₀
Reference law p₀Reference route probabilities π₀; every free corruption rate U(0,1), independent per stream (A2) or one shared draw per route (A1)
Route progressAbsolute decrease of the probe loss smoothed by an EMA with coefficient 0.9; standardized across routes; adaptive routing starts after the second probe
Route softmaxTemperature τₒ = 1; usage bonus 0.5/(1 + u), u = share of the actions drawn since the previous probe
Adaptive rate policyOne Beta per free route/stream (one per route for shared rates); initialized at Beta(1,1); r = 0.2 + 0.8v
Rate updateScore-function update, policy learning rate 0.05; reward-baseline EMA 0.9; log shape parameters clipped to [−1.5, 3]; ε = 10⁻⁴
Rate rewardPer-route mean of loss before minus loss after the optimizer step, on the identical corrupted batch with the pre-forward dropout RNG restored; per-record losses average to the training loss
HuBMAP / BioMed target weightsw ∝ (1 + γc + κe)·σ((c − b₀)/τw), c = max-class probability, e = 1 − p(gold); γ = 1, κ = 2, b₀ = 0.5, τw = 0.1; detached; normalized to mean 1 over corrupted targets
JUMP target weightsw = 1 + wβ·σ((cos − 0.5)/0.15), wβ = 1, on records with SMILES visible and both vectors hidden; normalized to mean 1
HuBMAP example samplingStrength 1; loss EMA 0.9; multiplier clip [0.5, 2]; ramp starting at 12% of training
Replay (B8 / B9)Every B1 update logs the policy its actions were drawn from. B8 permutes complete probe windows and keeps their internal order; B9 draws each action from the policy of a uniformly chosen B1 update

Table R6-B — JUMP training recipe, objective and selection

SettingUniDiscLLaDA2
Backbone / adaptationSmall: width 768, 12 layers, trained from scratch with ContIO (Tiny 384/6, Base 1024/16, Large 1536/24)LoRA rank 32, α = 64, dropout 0.05 (attention and shared-expert projections); reserved embedding/head rows and ContIO trainable
OptimizerAdamW, β = (0.9, 0.95); gradient clip 1Same
Learning ratesBackbone 3×10⁻⁴; ContIO 10⁻³LoRA and reserved rows 10⁻⁴; ContIO 10⁻³
Weight decayBackbone 0.01; ContIO 0LoRA 0.01; reserved rows and ContIO 0
Schedule200-update warm-up, then cosine on the optimizer-update counter150-update warm-up, then cosine
Batch64 (one GPU)Global 64 = 4 GPUs × 16
Budget / checkpoints36k updates / every 4k10k updates / every 2k; longer panel 18k
Curriculum objectiveEq. 3 per record (1/max(r, 10⁻³) per hidden stream; vector loss (1 − cos) + 0.5·MSE); InfoNCE weight 0.1, temperature 0.07; bioactivity focal γ = 2Same
Endpoint terms (every arm, every update)T7: SMILES → bioactivity, class-balanced CE (positive weight 2), weight 1; T8: SMILES → morphology regression + InfoNCE 0.1/0.07, weight 1Same
Plain-denoising panelInfoNCE and focal modulation removedSame
Training probe2,048 training compounds drawn with seed 0; refresh every 500 updatesSame
Reference corruptionπ₀(joint) = 1; independent U(0,1) per-modality rates; whole-modality masking; label hidden with probability 0.5 on direct/joint; guard: ≥ 1 visible input, ≥ 1 hidden targetSame
Route probesdirect: SMILES visible, morphology/structure hidden; s2m: structure visible, morphology/SMILES hidden; m2s: morphology visible, structure/SMILES hidden; joint: rate 0.5 per streamSame
SelectionValidation only: composite (mean of T1, T3sm, T2mc Hits@10 and T7 pAUROC) and per-endpoint checkpoints; T8 selected by validation T7Same
AuthorsOpenReview ↗

Response to Reviewer yNzh

Table R6-C — HuBMAP-MR training and evaluation recipe

SettingLLaDA2UniDisc
Backbone / adapterLoRA on the last 2 layers, rank 8, α = 16, dropout 0.05147M parameters; 12 layers, width 768; trained from scratch
Spatial interface1,024 image tokens; 16×16 (256) or 32×32 (1,024) mask tokens; organ prompt; residual read-out width 128MagViT tokens of 256-pixel tiles; text region 64 + G² tokens holding the organ prompt and G×G mask tokens (G = 16 or 32)
Read-out featuresImage positions of layers 4/12/20 plus projected mask-position statesImage and mask-query positions of blocks 4/8/12
OptimizerAdamW, β = (0.9, 0.95), weight decay 10⁻³, gradient clip 1AdamW, β = (0.9, 0.95), weight decay 0.05, gradient clip 1
Learning rates / scheduleLoRA 2×10⁻⁵; read-out 3×10⁻⁴; 100-update warm-up, then cosineBackbone 10⁻⁴; read-out 3×10⁻⁴; 200-update warm-up, then cosine
Updates / batch12k; batch 2 × accumulation 412k; batch 8
Mask read-out lossesBCE positive weight 3; Dice 2; prevalence 2; boundary 0.5; corrupted mask cells onlySame
Token CEEq. 3 with coefficients mask 0.05, organ 0.05, image 0.10Same
Reference route probabilitiesSegmentation 0.40; joint 0.30; image completion 0.15; organ completion 0.15Same
Canonical probe ratesSegmentation mask 1; organ completion organ 1; image completion image 0.5; joint all streams 0.5Same
Training probe64 iid_train tiles per route / every 200 updates256 tiles per route / every 500 updates
SelectionCheckpoints every 1k; iid_cal mean native-image Dice; threshold sweep 0.01–0.99Same
InferenceOne forward with the whole mask hidden; image and organ visible; native-resolution reconstructionSame

Table R6-D — BioMed recipe

SettingValue
Backbone / trainable modulesLLaDA2 LoRA rank 32, α = 64, dropout 0.05 on attention qkv/dense, shared-expert gate/up/down, lm_head and word embeddings
OptimizerAdamW, β = (0.9, 0.95), weight decay 0.01
Learning rate / schedule2×10⁻⁵; 20-update warm-up, then cosine
Batch / hardwareBatch 2 per GPU, accumulation 8; 8× NVIDIA B200
Training horizons1× = 30k steps; 2× = 60k steps; checkpoints every 5k
Shared initial phaseSegmentation route only for the first 40% of training in every arm, with reference rates
Reference patternsSegmentation 0.25; text reconstruction 0.20; joint mask/text 0.55; image always visible
LossEq. 3: 1/max(r, 10⁻³) per stream; two-class mask CE with foreground weight 3; text CE; SoftDice coefficient 5
Training probe64 training records per pattern; every 200 optimizer updates; fixed corruption seed 0
Canonical probe ratesSegmentation mask 1; text reconstruction text 0.5; joint mask/text 0.5/0.5
Target weightsTable R6-A; mask cells use the two-class posterior, text tokens the full-vocabulary posterior; applied to mask CE, text CE and Dice
Data splits16×16 grid, cell foreground at ≥ 25% occupancy; held-out adapter trains on the non-held-out sources' IID training records; held-out sources split into target-calibration (20%) and test records
EvaluationAll 256 mask cells hidden; threshold (0.05–0.95) and checkpoint chosen on calibration data: source = iid_cal, target = held-out-source calibration records; held-out adapter: 1,397 held-out and 4,096 IID test records; IID adapter: 11,229 records; global Dice, image Dice, AUROC
Reviewer_x9YqOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper aims to address the problem of diffusion learning across heterogeneous multimodal data. It aims to answer the question of how a learner should distribute its optimization as training progresses in this context.

Critical strengths

The authors use a novel approach to address the problem of multimodal diffusion for biological datasets. It seems that their results may have significance for both researchers interested in multimodal biological datasets, and those interested in multimodal diffusion broadly.

Critical weaknesses

Structurally, it would be better if the authors outlined the desiderata for a multimodal diffusion model for biology more clearly at the beginning of the paper, so that as the approach is being described they could link back to it. Right now some of the desiderata seem to be first well defined in the theoretical analysis section.

This paper could use a close proof read, to tight up language and improve clarity. Nits:

  • The introduction starts with almost the same clause as the abstract - it’s redundant.
  • Some in-text citations are missing a space between the prior word e.x. “information(Zhao et al., 2025).”
  • Some text in Figure 1 is way too small
Questions for authors

No questions that would materially affect my assessment.

1 comment
AuthorsOpenReview ↗

Response to Reviewer x9Yq

Dear Reviewer x9Yq,

Thank you for your constructive suggestions and insightful comments! We thank the reviewer for the presentation suggestions. The revised framing connects the training design to four explicit requirements and to the evidence testing each one. We also provide further clarifications and refinements in the general response and paper revision.

W1: State the desiderata at the beginning and link back to them

It would be better if the authors outlined the desiderata for a multimodal diffusion model for biology more clearly at the beginning of the paper, so that as the approach is being described they could link back to it.

We appreciate this point. We will introduce the training requirements before the contributions:

A training procedure for multimodal biological diffusion should cover asymmetric prediction problems, adapt allocation to the learner, preserve reference exposure while concentrating effort, and support heterogeneous supervised denoising objectives. Independent corruption supplies directional coverage; learned allocation adapts training effort; reference mixing maintains exposure; and example/target factors distribute learning within the supported problems.

Each requirement has an empirical counterpart: A1/A2 tests support; A2/A4 and B1/B4 test adaptation; the reference mixture and controller snapshots document exposure; and the objective panel examines auxiliary-loss sensitivity. The five-choice audit measures how useful training blocks vary with learner state. We will use these links to connect the introduction, method and experiments.

W2: Proofreading and specific nits

The introduction starts with almost the same clause as the abstract; some in-text citations are missing a space; some text in Figure 1 is way too small.

Thanks for the suggestions. We have revised the introduction: “Biological datasets connect perturbations, molecular structure, morphology and spatial annotations through prediction problems with different directions and difficulty.” In the manuscript revision we correct inconsistent citation spacing, and reduce figure-label density and enlarge Figure 1 text at its final display size.

We sincerely thank you for your time! Hope we have addressed your concerns through practical efforts and shown the contributions and significance of our work. We look forward to your reply and further discussions, thanks!

Sincerely,

Authors

Reviewer_2VHnOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MaDAC, a curriculum for multimodal diffusion models. It adjusts which prediction tasks the model practices and how much each modality is corrupted, then reweights individual training targets. The motivation is that different tasks may be useful at different stages of learning. Experiments cover compound–phenotype prediction and biomedical segmentation, with the strongest gains on BioMed.

Critical strengths
  • The motivation is quite sensible as a shared corruption rate restricts the training tasks a model encounters. Proposition 3.2 explains why independently controlling modalities can expose useful cases that shared-rate training rarely visits.
  • The paper is well written and easy to follow
Critical weaknesses
  • MaDAC changes both the available corruption combinations and how they are selected. A non-adaptive sampler with independent per-modality rates could already address the main motivation. I could not find a matched comparison that clearly separates these benefits.
  • Theorem 3.3 establishes an advantage when task usefulness can be estimated accurately enough. The paper does not establish that its actual progress signals meet this condition. This limits how directly the theorem supports the proposed method.
  • BioMed targets are reduced to a 16×16 grid, which can remove small structures and boundary detail. The main results should state the evaluation resolution explicitly and clarify what the gains imply for native-resolution segmentation.
  • For LLaDA2 in Table 1, structure-to-morphology retrieval improves, while the other four endpoints change little or even decline slightly. HuBMAP uses a fixed prediction route, so its gains provide less direct evidence for adaptive multimodal task selection.
Questions for authors

Please see the weaknesses above

5 comments
AuthorsOpenReview ↗

Response to Reviewer 2VHn

Dear Reviewer 2VHn,

Thank you for your constructive suggestions and insightful comments! We thank the reviewer for the questions about support, adaptive selection, estimator quality and spatial fidelity. Our revised experiments directly separate these mechanisms and report their effects by task and endpoint. We also provide further clarifications and refinements in the general response and paper revision.

W1. Independent per-modality support versus adaptive selection

MaDAC changes both the available corruption combinations and how they are selected. ... I could not find a matched comparison that clearly separates these benefits.

Thank you for raising this point. We provide a set of ablation studies around this question, as specified in General Response above. We also further clarify in the anonymous code and revised paper.

MaDAC improves over A2, the requested fixed independent control. On BioMed, B1 − A2 improves held-out global Dice by +0.046 / +0.085 at 1× and +0.084 / +0.018 at 2×, in source/target-calibration order. Both 1× comparisons have positive seed agreement, CIs excluding zero and Holm-adjusted improves calls. The separately trained IID adapter gains +0.044 [95% CI +0.042, +0.046] (G.7).

The A1–A4 factorial comparison attributes BioMed's benefit to adapting separate rates. A4 − A2 gives +0.112 / +0.096 at 1×, and A4 − A3 gives +0.051 / +0.065. Separate fixed rates (A2 − A1) show no consistent held-out gain. Thus, separate rates help when adapted.

B1 − B4 tests learned allocation with the other factors retained. BioMed AUROC improves by about +0.035 under both calibration rules, and separately trained IID global Dice improves by +0.021. The BioMed trajectory remains positive against A2 at every checkpoint through 30k and at most checkpoints through 60k (R7-T-BioMed).

JUMP B1 exceeds A2 in all three reported retrieval comparisons on both backbones; B1 − B4 resolves three retrieval endpoints on UniDisc and SMILES→morphology/bioactivity on LLaDA2 (G.5). HuBMAP-MR B1 exceeds B4 in mean image Dice on both backbones (G.6).

W2. Does the actual progress signal meet the condition of Theorem 3.3?

Theorem 3.3 establishes an advantage when task usefulness can be estimated accurately enough. The paper does not establish that its actual progress signals meet this condition.

This is an interesting question and we try to study empirically. We conduct a five-choice HuBMAP audit which yields a positive empirical covered bound and a positive average utility advantage over the best fixed block policy. It evaluates the theorem's finite-choice mechanism over 500-update interventions.

The audit evaluates six HuBMAP UniDisc checkpoints, 2k, 6k and 10k updates for each of the training seeds, equally weighted. Each candidate runs for 500 updates from the same restored model, read-out, optimizer, controller and sampling state, with four branch seeds. Controller adaptation is frozen within each branch; parent rate policies, sampling and weighting remain active. Utility is canonical segmentation-loss reduction on 61 iid_cal images.

Table R4-P — five candidate block interventions

Block choiceKeyIntervention over 500 updates
SegmentationsegUse the image and organ label to predict the mask.
Joint completionjointRecover the partly corrupted image, mask, and organ jointly.
Image completionrec_iUse the mask and organ label to recover the image.
Organ completionrec_oUse the image and mask to predict the organ label.
Reference mixturerefSample the four routes with probabilities 0.40, 0.30, 0.15, and 0.15 in table order.

The covered block policy is q = (1−ρ)π + ρp₀ᴮ, with ρ = 0.20 and p₀ᴮ = (0.40, 0.30, 0.15, 0.15, 0). The comparator uses one adjustable distribution across all six states with the same blocks, state weights and coverage. Its optimal adjustable mass is on the reference block, which has the highest mean empirical utility.

Table R4-A — state-by-state five-choice audit

CheckpointEmpirical oracleController's top choiceAgreementUtility advantage over the best fixed policy ↑
2kSegmentationSegmentationYes+0.030000
6kImage completionImage completionYes+0.037040
10kReference mixtureReference mixtureYes−0.013440

The controller selects the empirical winner and achieves +0.008228 mean utility advantage. For block utility Q₅₀₀(s,b), opportunity is Γ₅₀₀ = Eμ max_b Q₅₀₀(s,b) − max_b Eμ Q₅₀₀(s,b). With a uniform candidate-error bound ε and selection gap ζ, the finite-choice argument gives

V₅₀₀(q) − Vstat,₅₀₀^ρ ≥ (1−ρ)[Γ₅₀₀ − 2Eμε₅₀₀ − Eμζ₅₀₀].
AuthorsOpenReview ↗

Response to Reviewer 2VHn

The sufficient condition is the pooled inequality Γ₅₀₀ > 2Eμε₅₀₀ + Eμζ₅₀₀, which compares the opportunity from adapting across states with estimation and selection costs.

With calculation, opportunity Γ = 0.0295 exceeds the combined estimation/selection cost of 0.026103 (twice the mean candidate error is 0.008 and the mean selection gap is 0.018). After coverage, the empirical lower bound is +0.002717, below the directly evaluated advantage of +0.008228. The bound applies algebraically to empirical branch means on six saved states and 61 calibration images. Theorem 3.3 defines one-update utility. Here each intervention controls an entire 500-update block, so the audit tests the corresponding finite-block argument. Thus it supports the proposed mechanism; the most useful block changes with learner state, and the five-choice controller captures enough of that opportunity to achieve a positive empirical bound and average utility advantage at the tested horizon.

While practical conditions do not identically mimic theoretical conditions, empirically we see our progress signals agreeing with the proposed theoretical mechanism.

W3: The 16×16 output grid, small structures and native-resolution claims

BioMed targets are reduced to a 16×16 grid ... The main results should state the evaluation resolution explicitly and clarify what the gains imply for native-resolution segmentation.

We first provide several additional ablation studies around resolution to study how resolution affect performances. Then we clarify our choice of resolution due to compute resource constraints.

We experiment with HuBMAP, and separates output resolution from scoring resolution: predictions use the stated mask grid, followed by reconstruction and a common native-resolution evaluator. This tests segmentation quality under a coarse output interface. BioMed's current exports support global/image Dice and include diagnostic test1024 oracle fields.

The HuBMAP-MR grid and spatial studies quantify both prediction quality and the detail retained by the output interface.

Table R5-A / source 5A — HuBMAP native-scored quality by mask grid

BackboneMask gridArmOOD image Dice ↑OOD macro Dice ↑OOD empty % ↓
LLaDA216×16A2 · Uniform0.336 ± 0.0060.343 ± 0.01423.661 ± 2.210
LLaDA216×16B1 · MaDAC (Algorithm 1)0.458 ± 0.0300.474 ± 0.0220.000 ± 0.000
LLaDA232×32A2 · Uniform0.284 ± 0.0220.294 ± 0.04043.192 ± 0.158
LLaDA232×32B1 · MaDAC (Algorithm 1)0.425 ± 0.0010.432 ± 0.0100.000 ± 0.000
UniDisc16×16A2 · Uniform0.307 ± 0.0270.338 ± 0.03133.371 ± 13.732
UniDisc16×16B1 · MaDAC (Algorithm 1)0.399 ± 0.0100.419 ± 0.0080.335 ± 0.158
UniDisc32×32A2 · Uniform0.269 ± 0.0180.295 ± 0.04447.545 ± 14.205
UniDisc32×32B1 · MaDAC (Algorithm 1)0.370 ± 0.0450.382 ± 0.0190.781 ± 0.158

B1 retains higher mean image Dice at both grids and on both backbones. Its gap over A2 increases from 0.122 to 0.141 on LLaDA2 and from 0.092 to 0.101 on UniDisc. Absolute Dice decreases at 32×32 for both arms: LLaDA2 A2/B1 changes from 0.336/0.458 to 0.284/0.425; UniDisc changes from 0.307/0.399 to 0.269/0.370. Thus finer output capacity and learned prediction quality have separate effects at the tested budget. These gap changes are descriptive. The 32×32 UniDisc empty rates follow the table entries, 47.545% and 0.781%.

The sparse-foreground subset comprises the lowest quartile by ground-truth foreground fraction: the same 112 source-held-out images per seed across the reported comparisons.

Table R5-B / source 5B — spatial diagnostics and grid reconstruction references

BackboneGridArmSmall-gold-area image Dice ↑Subset n/seedGrid reconstruction image Dice ↑
LLaDA216×16A2 · Uniform0.269 ± 0.0031120.907
LLaDA216×16B1 · MaDAC (Algorithm 1)0.327 ± 0.1131120.907
LLaDA232×32A2 · Uniform0.173 ± 0.0071120.944
LLaDA232×32B1 · MaDAC (Algorithm 1)0.367 ± 0.0221120.944
UniDisc16×16A2 · Uniform0.113 ± 0.0791120.907
UniDisc16×16B1 · MaDAC (Algorithm 1)0.210 ± 0.0111120.907
AuthorsOpenReview ↗

Response to Reviewer 2VHn

Table R5-B.1 — interpretation and coverage of the spatial diagnostics

Framework elementHow it is representedWhat it resolves
Output grid versus scoring resolution16×16 and 32×32 label the output interface; Dice is calculated after native-resolution reconstructionDistinguishes native-resolution scoring from native-detail output
Sparse foregroundSmall-gold-area Dice and 112 subset images per seedMeasures overlap for images with low foreground occupancy
Discretization referenceNative Dice of ground truth reconstructed through each gridQuantifies the overlap retained by the specified downsampling and reconstruction procedure
Overall and subset qualityOverall image/macro Dice and empty rate in 5A, subset Dice in 5BShows whether the observed pattern holds in sparse-foreground images where matching rows are available
Method identityA2 and B1 have separate rowsIdentifies which controller each spatial result evaluates
Spatial coverageOccupancy and reconstruction overlapBoundary fidelity and individual component recall require dedicated metrics
Evidence boundaryThe occupancy subset and reconstruction Dice do not measure boundaries or individual small componentsMakes the partial answer to 2VHn W3 explicit; no boundary fidelity or component preservation is established by these columns alone

Ground-truth reconstruction Dice increases from 0.907 to 0.944 at the finer grid, quantifying the additional detail retained by the interface. At 16×16, B1 also exceeds A2 in mean sparse-foreground Dice on both backbones.

Table R5-C / source 5C — output-interface cost

LLaDA2 adapter benchmarks use one H200. Coarse-grid costs are reported measurements; native-interface timings are extrapolated. Wall-time estimates correspond to about 7.5k BioMed updates and 12k HuBMAP updates.

TaskMask gridMask tokensTime/update (s unless stated)Peak memorySource-stated wall time
BioMed16×1625615.236.8 GB32 h
BioMed32×32102422.546.4 GB47 h
BioMed64×64409651.884.5 GB108 h
BioMed128×12816384OOMOOM (>141 GB)
BioMedNative1,048,576≈2.9 h≈2.5 years
HuBMAP16×162563.134.6 GB10 h
HuBMAP32×3210243.336.3 GB11 h
HuBMAP64×6440964.443.1 GB15 h
HuBMAP128×128163849.970.2 GB33 h
HuBMAPNative262,144≈8.7 min≈73 days

The cost panel quantifies the spatial interface's compute trade-off. HuBMAP completes its 128×128 benchmark at 9.9 s/update and 70.2 GB; BioMed reaches OOM at that grid. Native-grid rows provide projections. Quality conclusions use the trained grids in R5-A, and compute conclusions use the benchmarked interfaces in R5-C.

W4: LLaDA2 on JUMP, and HuBMAP's fixed prediction route

For LLaDA2 in Table 1, structure-to-morphology retrieval improves, while the other four endpoints change little or even decline slightly. HuBMAP uses a fixed prediction route, so its gains provide less direct evidence for adaptive multimodal task selection.

We have conducted the experiments with multiple seeds to carefully re-validate our method, as the General Response.

The updated JUMP B1 comparison has higher mean scores than A2 on all five endpoints with both backbones. G.5 presents the endpoint tables (R2-U/R2-L) and component comparisons. The reported B1–A2 categories demonstrated all retrieval endpoints as positive.

HuBMAP-MR now includes allocation across four prediction routes. G.6 introduces the corruptible-mask interface and full ablations. B1 exceeds fixed-allocation B4 in mean image Dice on UniDisc (0.399 versus 0.355) and LLaDA2 (0.458 versus 0.410), with the example/target factors retained. Segmentation-only R2/R3 controls additionally test the choice of route set. The reconstructed setting therefore supplies a direct test of learned allocation within a common multi-route segmentation task.

We sincerely thank you for your time! Hope we have addressed your concerns through practical efforts and shown the contributions and significance of our work. We look forward to your reply and further discussions, thanks!

Sincerely,

Authors

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Response

We thank all reviewers for their valuable feedback and discuss a few points in this general response.

We have clarified Algorithm 1 and rerun controlled ablations on JUMP, BioMed and HuBMAP-MR. Throughout this response, A2 represents Uniform, the nonadaptive independent-stationary control, A1 is the shared-rate control, and B1 represents MaDAC. The revised paper will include the clarified algorithm and controls. HuBMAP-MR directly tests allocation across prediction routes. The mask is a corruptible token stream alongside image and organ streams. Training selects segmentation, joint completion, image completion and organ completion. Mask read-out supervision applies to corrupted mask cells, and masked organ information is hidden from the read-out embedding. Matched arms share the read-out/loss interface and train the adapter/backbone from the first update. Inference hides the entire mask and uses a common native-resolution reconstruction and evaluator.

The revised experiments clarifies our contributions: gains over the requested nonadaptive independent control, direct attribution of the learned allocation and its supporting components, and evidence that the gains persist across the tested model sizes and training budgets. The ablations identify how these contributions depend on the task, backbone and endpoint.

G.1 Clarified findings

BioMed: Algorithm 1 improves over fixed, separate per-modality rates. B1 exceeds A2 in held-out global Dice under both calibration rules: +0.046 / +0.085 at 1× and +0.084 / +0.018 at 2×, in source/target-calibration order. The separately trained IID adapter gains +0.044 [95% CI +0.042, +0.046]. The A1–A4 factorial comparison identifies rate adaptation as the key contributor: separate rates help when adapted. B1–B4 further shows improved AUROC under both calibration rules and improved IID Dice (G.7).

JUMP: B1 achieves the highest mean on all five endpoints with both backbones. The reported B1–A2 comparisons resolve all three retrieval endpoints as positive. The model-size and training-budget studies show that the retrieval advantage persists across the tested regimes (G.5 and R7).

HuBMAP-MR: learned allocation improves segmentation with example sampling and target weighting held fixed. B1 raises source-held-out mean image Dice from A2's 0.307 ± 0.027 to 0.399 ± 0.010 on UniDisc and from 0.336 ± 0.006 to 0.458 ± 0.030 on LLaDA2. B1 also outperforms fixed-allocation B4 on both backbones (G.6).

G.2 Shared experimental design and supporting evidence

The controls separate the reviewers' attribution questions:

ComparisonQuestion studied
A2 − A1Does fixed, separate per-modality corruption improve over a shared rate?
A4 − A2Does adapting separate rates help?
A4 − A3Do separate rates help when both policies adapt?
B1 − A2Does full MaDAC improve over the requested nonadaptive independent control?
B1 − B4Does learned allocation help with the other MaDAC factors retained?

A1 uses a shared stationary rate; A2 (Uniform) uses independent stationary rates; A3 adapts a shared rate; A4 adapts independent rates. B1 runs Algorithm 1. B4 fixes routes and rates at the reference law while retaining target weighting and, on HuBMAP-MR, example sampling. B2/B3 remove learned routes/rates separately; B5/B6 remove target weighting/example sampling; B7 removes both. B8/B9 test allocation timing where available.

G.5–G.7 present JUMP, HuBMAP-MR and BioMed results, respectively. BioMed retains the A1–A4 rate design and B1/B4 comparisons needed for the BioMed attribution questions. JUMP and HuBMAP-MR additionally report component and replay controls. HuBMAP-MR uses a common read-out/loss interface, training budget and learning-rate warm-up across matched arms, with the adapter/backbone trained from the first update. Its segmentation-only R2/R3 controls test the route set.

The five-choice HuBMAP audit tests adaptive utility (R4). Resolution, compute, model-size, training-budget and data/noise panels characterize the tested operating regimes (R5 and R7). R6 records training recipes, controller parameters and data roles.

G.3 Method identity, endpoints and uncertainty

Method identity. MaDAC(B1) learns route allocation and bounded rate policies using absolute route progress and the same-batch improvement reward, with reference mass ρ = 0.20 and target weighting. HuBMAP also uses example sampling. BioMed and JUMP use uniform example sampling. A2 holds independent rates and reference routes stationary, with target weighting off. B4 fixes allocation while retaining the other B1 factors; B5 removes target weighting.

AuthorsOpenReview ↗

General Response

Endpoints. BioMed reports pooled/global Dice, equal-record image Dice and full AUROC (area under the receiver operating characteristic curve). HuBMAP's primary endpoint is source-held-out native-scored image Dice, accompanied by macro-organ Dice and empty-prediction rate. IID denotes independent and identically distributed evaluation; HO and OOD denote source-held-out evaluation here. JUMP reports SMILES→morphology, structure→morphology and morphology→compound Hits@10, followed by bioactivity and balanced mechanism-of-action (MoA) partial AUROC.

Uncertainty. Main scores are means ± SD over training seeds. Resolution/scale panels retain their reported SDs, and the five-choice audit reports point estimates.

G.4 How we use comparisons

Absolute scores provide the main evidence; secondary contrasts address the reviewers' attribution questions. The JUMP, HuBMAP-MR and selected BioMed panels follow in G.5, G.6 and G.7, respectively. We also discuss more in the responses to each reviewer, while updated paper pdf provides further information.

G.5 JUMP: revised results and component comparisons

The JUMP comparison leads on all five mean endpoint scores with both backbones.

Table R2-U — JUMP UniDisc merged results

IDArmSMILES→morph H@10 ↑Structure→morph H@10 ↑Morph→compound H@10 ↑Bioactivity pAUROC ↑MoA pAUROC ↑Coverage
A1Shared-rate0.501 ± 0.0000.607 ± 0.0020.633 ± 0.0010.679 ± 0.0050.752 ± 0.002
A2Independent-stationary (matched rerun of Uniform law)0.495 ± 0.0080.606 ± 0.0040.636 ± 0.0010.682 ± 0.0050.754 ± 0.001
A3Shared-rate, adapted0.537 ± 0.0090.604 ± 0.0040.633 ± 0.0010.681 ± 0.0050.753 ± 0.002
A4Independent rates adapted alone0.539 ± 0.0040.615 ± 0.0030.635 ± 0.0010.681 ± 0.0040.751 ± 0.001
B1MaDAC (Algorithm 1)0.571 ± 0.0020.628 ± 0.0040.646 ± 0.0040.686 ± 0.0030.758 ± 0.003
B2MaDAC − learned routes0.543 ± 0.0010.614 ± 0.0030.636 ± 0.0040.681 ± 0.0050.756 ± 0.001
B3MaDAC − learned rates0.549 ± 0.0060.619 ± 0.0010.638 ± 0.0010.684 ± 0.0000.753 ± 0.000
B4MaDAC − learned allocation0.551 ± 0.0050.618 ± 0.0020.637 ± 0.0060.683 ± 0.0050.755 ± 0.000
B5MaDAC − target weighting0.539 ± 0.0010.615 ± 0.0040.635 ± 0.0010.679 ± 0.0050.752 ± 0.001
B7MaDAC − sampler − target weighting0.539 ± 0.0010.615 ± 0.0040.635 ± 0.0010.679 ± 0.0050.752 ± 0.001Same experiment as B5; example sampling is uniform.

Table R2-L — JUMP LLaDA2 merged results

IDArmSMILES→morph H@10 ↑Structure→morph H@10 ↑Morph→compound H@10 ↑Bioactivity pAUROC ↑MoA pAUROC ↑Coverage
A1Shared-rate0.539 ± 0.0120.615 ± 0.0030.636 ± 0.0010.673 ± 0.0010.748 ± 0.001
A2Independent-stationary (matched rerun of Uniform law)0.553 ± 0.0120.618 ± 0.0030.638 ± 0.0020.681 ± 0.0000.749 ± 0.000
A4Independent rates adapted alone0.553 ± 0.0120.621 ± 0.0010.639 ± 0.0020.682 ± 0.0000.751 ± 0.001
B1MaDAC (Algorithm 1)0.565 ± 0.0120.625 ± 0.0030.648 ± 0.0010.685 ± 0.0000.752 ± 0.000
B2MaDAC − learned routes0.551 ± 0.0120.623 ± 0.0030.641 ± 0.0020.683 ± 0.0000.750 ± 0.000
B3MaDAC − learned rates0.549 ± 0.0120.625 ± 0.0010.641 ± 0.0020.683 ± 0.0000.752 ± 0.000
B4MaDAC − learned allocation0.547 ± 0.0100.624 ± 0.0020.643 ± 0.0020.680 ± 0.0030.749 ± 0.002
B5MaDAC − target weighting0.550 ± 0.0120.623 ± 0.0030.644 ± 0.0010.685 ± 0.0000.751 ± 0.001
B7MaDAC − sampler − target weighting0.550 ± 0.0120.623 ± 0.0030.644 ± 0.0010.685 ± 0.0000.751 ± 0.001Same experiment as B5; example sampling is uniform.

JUMP endpoint comparisons and supporting panels

The reported B1–A2 categories demonstrate all three retrieval endpoints as positive on both backbones. B1–B4 resolves three retrieval endpoints on UniDisc and SMILES→morphology/bioactivity on LLaDA2. These comparisons directly support the contribution of learned allocation, with endpoint-specific resolution. The panel pairs plain-denoising and default-objective results.

AuthorsOpenReview ↗

General Response

Table R2-b — supplied objective-robustness panel

IDArmTraining objectiveSMILES→morph H@10 ↑Structure→morph H@10 ↑Morph→compound H@10 ↑Bioactivity pAUROC ↑MoA pAUROC ↑
A1Shared-ratePlain denoising0.489 ± 0.0040.606 ± 0.0000.637 ± 0.0010.672 ± 0.0040.745 ± 0.006
A1Shared-rateDefault objective0.501 ± 0.0000.607 ± 0.0020.633 ± 0.0010.679 ± 0.0050.752 ± 0.002
A2Independent-stationary (matched rerun of Uniform law)Plain denoising0.490 ± 0.0010.606 ± 0.0020.634 ± 0.0010.678 ± 0.0050.748 ± 0.007
A2Independent-stationary (matched rerun of Uniform law)Default objective0.495 ± 0.0080.606 ± 0.0040.636 ± 0.0010.682 ± 0.0050.754 ± 0.001
B1MaDAC (Algorithm 1)Plain denoising0.559 ± 0.0010.625 ± 0.0060.639 ± 0.0020.682 ± 0.0030.750 ± 0.006
B1MaDAC (Algorithm 1)Default objective0.571 ± 0.0020.628 ± 0.0040.646 ± 0.0040.686 ± 0.0030.758 ± 0.003

Plain-denoising B1 outperforms A2 in mean score on all five endpoints.

G.6 HuBMAP-MR: revised results and component ablations

HuBMAP-MR makes route allocation testable within segmentation. Its corruptible mask stream supports four training routes under a common read-out/loss interface. All matched studies use the same training phases, full-mask hiding at inference and native-resolution evaluator.

Table R3-U / source H-U — HuBMAP-MR UniDisc

IDArmSrc-HO image Dice ↑Src-HO macro Dice ↑Src-HO empty % ↓
A1Shared-rate0.285 ± 0.0020.313 ± 0.00239.397 ± 5.524
A2Uniform (independent-stationary)0.307 ± 0.0270.338 ± 0.03133.371 ± 13.732
A4Independent rates adapted alone0.317 ± 0.0340.349 ± 0.03828.795 ± 8.523
B1MaDAC (Algorithm 1)0.399 ± 0.0100.419 ± 0.0080.335 ± 0.158
B2MaDAC − learned routes0.343 ± 0.0070.374 ± 0.0087.589 ± 2.210
B3MaDAC − learned rates0.360 ± 0.0980.379 ± 0.08813.616 ± 19.256
B4MaDAC − learned allocation0.355 ± 0.0040.387 ± 0.0023.795 ± 0.947
B5MaDAC − target weighting0.322 ± 0.0160.355 ± 0.01718.415 ± 0.158
B6MaDAC − example sampler0.290 ± 0.0100.317 ± 0.00943.304 ± 0.000
B7MaDAC − sampler − target weighting0.274 ± 0.0180.304 ± 0.01643.750 ± 0.631
B8MaDAC, own time-shuffled allocation0.320 ± 0.0870.347 ± 0.08010.826 ± 12.469
B9MaDAC, own matched-marginal allocation0.351 ± 0.0290.379 ± 0.0254.576 ± 0.789
R1Frozen backbone; read-out trained0.165 ± 0.0720.205 ± 0.06554.353 ± 15.310
R2Single segmentation route; no curriculum0.341 ± 0.0010.373 ± 0.00116.518 ± 2.210
R3Single-route MaDAC factors0.354 ± 0.0050.383 ± 0.00310.045 ± 0.000

Table R3-L / source H-L — HuBMAP-MR LLaDA2

IDArmSrc-HO image Dice ↑Src-HO macro Dice ↑Src-HO empty % ↓
A1Shared-rate0.349 ± 0.0110.371 ± 0.01723.884 ± 2.525
A2Uniform (independent-stationary)0.336 ± 0.0060.343 ± 0.01423.661 ± 2.210
A4Independent rates adapted alone0.356 ± 0.0090.373 ± 0.01120.313 ± 2.210
B1MaDAC (Algorithm 1)0.458 ± 0.0300.474 ± 0.0220.000 ± 0.000
B2MaDAC − learned routes0.439 ± 0.0200.460 ± 0.0230.112 ± 0.158
B3MaDAC − learned rates0.444 ± 0.0260.459 ± 0.0292.679 ± 1.578
B4MaDAC − learned allocation0.410 ± 0.0300.428 ± 0.0260.000 ± 0.000
B5MaDAC − target weighting0.395 ± 0.0090.409 ± 0.0080.670 ± 0.316
B6MaDAC − example sampler0.327 ± 0.0080.355 ± 0.00916.295 ± 0.947
B7MaDAC − sampler − target weighting0.327 ± 0.0090.335 ± 0.01219.420 ± 2.841
B8MaDAC, own time-shuffled allocation0.459 ± 0.0150.477 ± 0.0130.000 ± 0.000
R1Frozen backbone; read-out trained0.378 ± 0.0100.380 ± 0.00925.335 ± 0.789
R2Single segmentation route; no curriculum0.418 ± 0.0250.433 ± 0.03214.397 ± 13.732
R3Single-route MaDAC factors0.473 ± 0.0260.460 ± 0.0410.112 ± 0.158

Learned allocation improves the mean score with example/target factors fixed: B1 outperforms B4 on UniDisc (0.399 versus 0.355) and LLaDA2 (0.458 versus 0.410). Removing learned routes gives B2 scores of 0.343/0.439; removing learned rates gives B3 scores of 0.360/0.444. These controls distinguish learning the allocation within a route set from choosing the route set itself.

AuthorsOpenReview ↗

General Response

Example sampling and target weighting. Removing the example sampler (B6) gives mean image Dice of 0.290/0.327 on UniDisc/LLaDA2; removing target weighting (B5) gives 0.322/0.395. The allocation-only B7 gives 0.274/0.327, compared with A2's 0.307/0.336. Combined with B1–B4, these controls locate the gain in the combination of learned allocation and example/target factors.

G.7 BioMed: component ablations

B1 runs Algorithm 1 on BioMed and improves over A2's fixed, separate per-modality rates. The comparison holds under source and target calibration, at 1× and 2× training budgets, and for the separately trained IID adapter.

The 1× and 2× training horizons are 30k and 60k steps. The held-out adapter is evaluated on 1,397 held-out records and a secondary IID test of 4,096 records; the separately trained IID adapter is evaluated on 11,229 records. Scores are means ± SD across training seeds. Global Dice pools confusion counts; image Dice averages per-record Dice; AUROC measures ranking quality. Paired 95% CIs are conditional on the fitted runs.

B1 versus the fixed independent control (Table R1)

B1 − A2 improves held-out global Dice by +0.046 / +0.085 at 1× and +0.084 / +0.018 at 2×, in source/target-calibration order. At 2×, AUROC and the secondary IID test improve under both calibration rules. The separately trained IID adapter gains +0.044 [95% CI +0.042, +0.046].

Table R1.1 — Held-out adapter, 1×, source calibration

MethodGlobal Dice ↑AUROC ↑Image Dice ↑
A1 · Shared-rate0.422 ± 0.0070.913 ± 0.0210.419 ± 0.028
A2 · Uniform (independent-stationary)0.356 ± 0.0480.916 ± 0.0090.369 ± 0.068
A3 · Shared-rate adapted0.417 ± 0.0270.909 ± 0.0010.446 ± 0.018
A4 · Independent rates adapted alone0.468 ± 0.0070.949 ± 0.0020.526 ± 0.010
B1 · MaDAC (Algorithm 1)0.402 ± 0.0580.945 ± 0.0040.479 ± 0.044
B4 · MaDAC − learned allocation0.411 ± 0.0290.910 ± 0.0330.451 ± 0.064

Table R1.2 — Held-out adapter, 2×, source calibration

MethodGlobal Dice ↑AUROC ↑Image Dice ↑Secondary IID global Dice ↑
A2 · Uniform (independent-stationary)0.418 ± 0.0000.950 ± 0.0010.522 ± 0.0020.819 ± 0.007
A4 · Independent rates adapted alone0.496 ± 0.0030.955 ± 0.0090.585 ± 0.0050.841 ± 0.009
B1 · MaDAC (Algorithm 1)0.501 ± 0.0230.955 ± 0.0020.590 ± 0.0020.837 ± 0.001

Table R1-C — B1 − A2 global-Dice comparisons

| Adapter / budget | Calibration | Global Dice Δ [95% CI] | Seeds | | --- | --- | --- | --- | --- | | Held-out, 1× | Source | +0.046 [+0.035, +0.057] | All positive | | Held-out, 1× | Target | +0.085 [+0.073, +0.096] | All positive | | Held-out, 2× | Source | +0.084 [+0.068, +0.099] | All positive | | Held-out, 2× | Target | +0.018 [+0.005, +0.031] | All positive | | Separately trained IID, 1× | IID | +0.044 [+0.042, +0.046] | All positive |

G.8 What the revised comparisons establish

B1 improves over the requested fixed independent control. BioMed shows positive held-out global-Dice comparisons under both calibration rules and budgets, plus a positive separately trained IID comparison. JUMP resolves all three retrieval comparisons on both backbones. HuBMAP-MR shows higher mean image Dice on both backbones.

The A1–A4 design identifies what rate adaptation contributes. On held-out BioMed, A4 − A2 and A4 − A3 improve under both calibration rules. JUMP demonstrates A4 − A2 improvement on UniDisc's retrieval endpoints. HuBMAP-MR A4 exceeds A2 in mean image Dice.

B1 − B4 measures learned allocation with supporting factors fixed. On BioMed, this comparison improves AUROC under both calibration rules and improves separately trained IID Dice, and held-out global Dice. HuBMAP-MR B1 outperforms B4 in mean image Dice on both backbones. JUMP also demonstrates this on retrieval comparisons.

HuBMAP-MR component controls identify how allocation works with example and target factors. B5/B6 remove target weighting/example sampling, while B7 removes both. Together with B1 − B4, these controls show the contribution of allocation within the combined policy. UniDisc B1 also outperforms both replay controls in mean image Dice.

B1 implements Algorithm 1 across the revised task comparisons. On BioMed, A3/A4 retain reference route frequencies while B1 learns route allocation and rate policies. HuBMAP-MR supplies four prediction routes under a common segmentation interface.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.