PaperlayerView exact revision on arXiv

Source-linked AI summary

Lluna: Collaborative Unimodal Models for Missing-Modality Inference

Haoyue Bai, Yinyi Luo, Heqing Huang, Xiaoyuan Yi, James Evans, Jindong Wang

arXiv:2610.20526v1cs.MM

TL;DR

Prediction with missing modalities matters because streams, sensors, and clinical measurements may be unavailable, while collaboration gains can weaken or reverse as inputs disappear. LLUNA combines private prediction with progressive common-branch guidance; across eight datasets, it improves Mean-Missing over the strongest evaluated baseline on every task, averaging 4.26% relative improvement.

  • Problem

    Prediction with missing modalities matters because real-world streams and measurements may be unavailable, and collaborative gains can weaken or reverse as modalities disappear.

    Source: §1 · p. 1

  • Method

    LLUNA pairs direct private prediction from adapted features with a graph-based common residual branch, using PCC to align normalized common representations along single-modality removals.

    Source: §1 · p. 2, §3.4 · p. 5

  • Results

    Across eight datasets, LLUNA improves Mean-Missing over the strongest evaluated baseline on every task, averaging 4.26% relative improvement, with positive minimal-input mean gains on all eight tasks.

    Source: §1 · p. 2, §5 · p. 10

  • Takeaways & Limitations

    LLUNA maintains positive mean collaboration gains under minimal inputs across all eight tasks, supporting robust collaboration as available evidence decreases.

    Source: §5 · p. 10

    The evaluation removes entire modalities under known availability states; partial corruption, temporal misalignment, and naturally occurring missingness remain to be evaluated.

    Source: H Limitations · p. 56

Abstract

from arXiv · show

Pretrained unimodal models offer a flexible route to multimodal prediction through collaboration, but this collaboration can become less accurate than a surviving unimodal model when modalities are missing. We study this setting using fixed surviving-model anchors and find that collaborative gains can weaken or reverse as inputs disappear, even when the anchor's own input is unchanged. Our diagnostics further reveal reduced recovery of modality-specific information after aggregation and state-dependent collaborative representations. We propose Lluna, which combines a direct private prediction path from adapted unimodal features with a common graph-based collaboration branch that provides a residual correction. Progressive Collaboration Consistency organizes richer-to-poorer guidance along single-modality removals, aligning normalized common representations while task supervision adapts predictions to the available evidence. Across eight datasets spanning affective computing, scene understanding, and medical prediction, Lluna achieves the best evaluated average performance over non-full input states, improving Mean-Missing by 4.26\% on average relative to the strongest evaluated baseline. The fixed-anchor diagnostic also shows positive gains in mean collaboration under minimal inputs on all eight tasks. Code is available at https://github.com/AIFrontierLab/Lluna.

1 Introduction

LLUNA addresses the weakening or reversal of collaborative gains as modalities disappear by combining private prediction with common collaboration and progressive cross-state guidance. Across eight datasets, it improves Mean-Missing over the strongest evaluated baseline on every task.

1 Introduction

  • Collaboration gains can shrink or turn negative as modalities disappear, allowing an unchanged surviving unimodal model to outperform the collaborative predictor.Source: §1 · p. 1
  • Diagnostics find modality-specific signals less recoverable after aggregation and collaborative representations that vary across missing-modality states.The figure caption identifies private-signal recovery via linear-probe R2 and cosine distances between availability states as diagnostics.Source: §1 · p. 2, Figure 1 · p. 2
  • LLUNA combines a direct private prediction path from adapted unimodal features with a graph-based common collaboration branch that supplies a residual correction.Source: §1 · p. 2
  • Progressive Collaboration Consistency transfers guidance from richer to poorer states along single-modality removals, while a stability bound relates distant-state drift to accumulated local discrepancies.Source: §1 · pp. 2–3
  • Across eight datasets, LLUNA improves Mean-Missing over the strongest evaluated baseline on every task, averaging 4.26% relative improvement, with positive minimal-input mean collaboration gains on all eight tasks.The domain-averaged absolute decline in collaboration gain after two removals is 37.9% lower than with conventional collaboration.Source: §1 · pp. 2–3

2 Related Work

Prior work addresses multimodal collaboration, missing inputs, unified predictors, and multi-agent coordination through different modeling assumptions. LLUNA studies collaboration among independently pretrained unimodal models as modalities—and participating models—become unavailable.

Model collaboration

  • Collaboration methods combine predictions or representations through ensembling, learned interactions, shared embeddings, alignment, or graph-based aggregation.Source: §2 · p. 3

Multi-agent Systems

  • Unlike language-coordinated multi-agent systems, LLUNA collaborates over representations and needs neither conversational interfaces nor agentic behavior.Source: §2 · p. 3

Learning with missing modalities

  • Most missing-modality methods adapt a task-specific multimodal predictor, whereas LLUNA studies collaboration among independently pretrained unimodal models as inputs and participating models disappear.Source: §2 · p. 3

Omni and unified models

  • Omni and unified models integrate modalities into a common predictor or modeling framework for understanding and generation.Source: §2 · p. 3

3 Lluna

LLUNA combines private predictions from available unimodal models with graph-based common collaboration, then coordinates the common representations across progressively reduced input states. Its training also uses sampled removal chains, and the paper establishes a pathwise stability result for common representations.

3 Lluna

  • The setup uses labeled examples with M modalities and targets predictions for test inputs containing at most M−1 modalities.Source: §3 · p. 3

3.1 Overview of LLUNA

  • LLUNA combines direct private predictions from adapted unimodal features with a common collaboration branch that refines them through a compatibility graph.The unimodal models remain frozen during training; Progressive Collaboration Consistency guides common representations across poorer availability states.Source: §3.1 · pp. 3–4

3.2 Compatibility Graph Building

  • LLUNA builds a fixed compatibility graph from task-prediction agreement on training data, restricting and normalizing its weights for each available-model state.Positive self-loops preserve an incoming connection for an isolated available model.Source: §3.2 · p. 4

3.3 Residual Coordination of Common and Private Information

  • A private head predicts from adapted features, while graph message passing and attention pooling produce a common residual correction to that prediction.The correction is scaled by a learned scalar and added to the private prediction.Source: §3.3 · pp. 4–5

3.4 Training with Progressive Collaboration Consistency

  • PCC aligns unit-normalized common representations across adjacent states that differ by one removed modality, while task supervision adapts predictions to available evidence.Intermediate states can pass guidance along removal chains; training samples maximal chains and reuses shared states and transitions rather than evaluating every state at each step.Source: §3.4 · p. 5

3.5 Local-to-Global Representation Stability

  • Theorem 1 shows that controlling adjacent-state cosine discrepancies limits common-representation drift along a path between more distant availability states.The approach does not directly force distant evidence configurations to match.Source: §3.5 · p. 6

4 Experiments

Experiments compare LLUNA with collaboration and missing-modality baselines using a shared frozen model pool, evaluating Mean-Missing across incomplete input states on eight datasets in three domains. Ablations and diagnostics examine the prediction paths, consistency choices, and collaboration behavior as inputs disappear.

4.2 Main results

  • 4.26% relative improvement in Mean-Missing over the strongest evaluated baseline, with gains on all eight datasets.The evaluation averages task metrics equally over nonempty incomplete modality subsets.Source: Table 1 · p. 7, §4.2 · p. 7
  • LLUNA exceeds all ten controls on each of three tasks when baselines are trained with missing-input views.The strongest control scores 0.3546 versus 0.3421 MAE on SIMS, 32.64% versus 33.82% accuracy on WS, and 82.08% versus 84.17% balanced accuracy on ADNI.Source: §4.2 · pp. 7–8, Table 2 · p. 8

4.3 Ablation Study

  • Both prediction paths matter; compatibility beats uniform aggregation on seven tasks and ties ADNI, while adjacent PCC on common representations outperforms tested alternatives.Allowing consistency gradients to update both richer and poorer states worsens Mean-Missing on most tasks.Source: §4.3 · pp. 8–9
  • With PCC, common-state distances contract for both adjacent and more distant state pairs in the MELD, URF, and ADNI examples.The passage defines the reduction as the mean distance without PCC minus the mean distance with PCC, on matched test examples.Source: §4.3 · p. 9

4.4 Understanding Model Collaborations

  • 1.77%–27.02% relative MM improvement: nine-model pools beat three-model pools; LLUNA retains positive minimal-input gains on all tasks, while baseline gains are negative on six.Intermediate pool expansions are not uniformly beneficial; the baseline comparison uses a fixed unimodal reference for each modality.Source: §4.4 · pp. 9–10
  • Training-estimated compatibility graphs show sparse cross-modality bridges on MELD, stronger audio–visual coupling on WS, and dense connections on ADNI.The displayed graphs compare task-dependent interaction patterns.Source: §4.4 · p. 10

4.5 Training Efficiency and Hyperparameter Sensitivity

  • PCC budgets, widths, and depths are not uniformly better when increased; sparse chain training changes equal-task mean relative MM by −0.28% across eight tasks.On MELD, one sampled chain reduces comparisons per step from nine to two while improving MM by 0.49% relative to full training.Source: §4.5 · p. 10
  • 70.91% accuracy with 1.29M parameters on URF, versus 68.46% with 4.25M for LIMSSR; LLUNA uses 69.8% fewer parameters.The comparison excludes frozen backbones and separately fitted frozen heads from trainable parameter counts.Source: Figure 3 · p. 9, §4.5 · p. 10

5 Conclusion

LLUNA combines modality-specific prediction with residual common collaboration and improves Mean-Missing over the strongest evaluated baseline on every task. The appendix protocol clarifies how cross-dataset improvements and ranks are computed.

5 Conclusion

  • LLUNA improves Mean-Missing over the strongest evaluated baseline on every task, averaging 4.26% relative improvement across eight datasets.It also maintains positive mean collaboration gains under minimal inputs.Source: §5 · p. 10
  • Positive mean collaboration gains under minimal inputs support robust collaboration as available evidence decreases.The conclusion reports this diagnostic result across eight datasets.Source: §5 · p. 10

A.1 Selection and aggregation protocol

  • The reported relative improvement averages the per-dataset values equally across eight datasets, using unrounded values.The Avg. rank column averages each method’s ranks across datasets, with each dataset weighted equally.Source: §A.1 · p. 16

A.2 All methods under every availability state

The appendix provides complete method-by-state results for eight datasets; Table 4 covers SIMS and reports MAE for full and non-full input states.

A.2 All methods under every availability state

  • Tables 4–11 provide complete method-by-state results for all eight datasets; bold and underlining mark best and second-best means, not statistical significance.Ties are preserved when applying those rank marks.Source: §A.2 · p. 17
  • Table 4 reports SIMS MAE ×100, with trained-model five-run means and sample standard deviations and fixed-model single scores.It identifies F as full input and MM as the average over non-full states.Source: Table 4 · p. 17

MISSING-MODALITY METHODS

This section presents comparisons of missing-modality methods and Omni baselines, including benchmark-result rows and ADNI comparisons. It also describes task and input protocols, dataset-dependent model interaction graphs, and main-run training settings.

MISSING-MODALITY METHODS

  • The listed missing-modality methods include GIML, MARS, LIMSSR, SimMLM, ShaSpec, and Flex-MoE, with result rows reported across datasets.The supplied excerpts do not identify the column meaning for these values.Source: MISSING-MODALITY METHODS · p. 17

OMNI MODELS

  • The omni-model rows report results for Mean ens., Logit stack, Concat MLP, Attn. pool, Soft MoE, FoE, Set Trans., and CLARGA.The supplied table captions identify complete MOSI, MELD, URF, Daily, WS, and CirCor results and state that F is full input and MM averages non-full states, but the excerpts do not provide the column headers needed to map individual values to those states.Source: OMNI MODELS · p. 18, Table 5 · p. 18
  • ADNI’s Omni-model results include mean ensemble, logit stack, concat MLP, attention pooling, and soft-MoE comparators.The supplied result passages list scores without column labels, so they do not establish a winner or support attaching values to particular metrics or input states.Source: OMNI MODELS · p. 25

MISSING-MODALITY METHODS

  • ADNI’s missing-modality comparison includes GIML, MARS, LIMSSR, SimMLM, ShaSpec, and Flex-MoE.The supplied result passages list scores for these methods but do not provide column labels that support assigning individual values to specific metrics or availability states.Source: MISSING-MODALITY METHODS · p. 25

B.1 Tasks, splits, and availability states

  • CirCor evaluates murmur prediction across 14 incomplete auscultation-site combinations, while ADNI predicts conversion using MRI, PET, and clinical inputs, including clinical-absent states.CirCor removes each site's waveform and derived representations; ADNI aligns measurements within 90 days and defines converter outcomes using a first diagnosis within approximately 39 months, followed by confirmation.Source: §B.1 · p. 26

B.2 Frozen inputs and baseline interfaces

  • Reported links include twelve cross-modal links, with mean audio–visual coupling 0.77 versus subtitle–audio 0.21 and subtitle–visual 0.15; ADNI retains all 36 possible links, with cross-modality means of 0.76–0.80.Source: §B.2 · p. 29

B.3 Main-run optimization settings

  • Main-run training updates adapters and collaboration layers while keeping unimodal encoders fixed, and selects one checkpoint per run using validation Mean-Missing.Evaluation equally averages non-full states; the five runs share settings except for initialization, training, and chain-sampling seeds.Source: §B.3 · p. 29, Table 16 · p. 30

B.4 Missing-input training controls

The missing-input training evaluation uses SIMS, WorldSense, and ADNI with the main evaluation’s splits and test sets. It compares collaboration and Omni controls with LLUNA under missing-input training conditions.

B.4 Missing-input training controls

  • The evaluation uses SIMS, WorldSense, and ADNI with the same splits and test sets as the main evaluation, augmenting complete training examples with missing-input views.These tasks cover the paper’s three application domains.Source: §B.4 · p. 30
  • Collaboration controls share the frozen model pool and modality-level masking while retaining their defining fusion mechanisms and validation-Mean-Missing-selected checkpoints.I2MoE receives missing-input views through its own training procedure.Source: §B.4 · p. 30
  • Omni controls fit one shared task head on complete and missing-input views, while MM averages the six non-full test views equally.The tables report every test state for the ten controls.Source: §B.4 · pp. 30–31

C Additional Ablation Results

The ablations examine prediction paths, graph interactions, consistency choices, and hyperparameter sensitivity; pool analyses show that model count and composition both shape performance.

C Additional Ablation Results

  • Prediction-path controls compare private-only and common-only prediction while retaining the remaining within-task training settings.The private-only variant disables the common auxiliary loss and PCC; the common-only variant uses the unscaled common head.Source: C Additional Ablation Results · p. 32
  • The learned residual scale beats equal addition on six of eight tasks, while the preferred fixed coefficient varies across tasks.Concatenation changes both the prediction head and parameter count.Source: Table 20 · p. 32, C Additional Ablation Results · p. 32
  • The graph control replaces LLUNA’s compatibility graph with a complete active-node graph whose equal weights change connectivity when the compatibility graph is sparse.The comparison changes only the interaction mode.Source: C Additional Ablation Results · p. 32
  • Pair-selection ablations compare adjacent, Full-to-Minimal, and All-Pairs candidate sets, then sample up to Q unique single-removal pairs per step.For four independently maskable inputs, the candidate sets contain 28, 4, and 105 pairs, respectively.Source: C Additional Ablation Results · p. 32
  • PCC ablations compare detaching the richer-state target, targeting common or private representations, and averaging losses on both normalized representations.Removing stop-gradient gives both endpoints the consistency-loss gradient; task losses still supervise both states.Source: C Additional Ablation Results · pp. 32–33
  • The L9 design varies PCC strength, private/common width ratio, and graph depth, averaging three configurations per factor level without separately estimating interactions.Width and depth changes also change model capacity.Source: C Additional Ablation Results · p. 34

D.1 Model-Pool and Parameter Analyses

  • Nine-model pools give the best observed Mean-Missing performance on all eight tasks, while intermediate expansion effects vary by task.Daily rises from 35.59% to 37.85% and 38.42%; ADNI changes from 82.71% to 76.67% and 84.17%.Source: §D.1 · p. 37, Table 26 · p. 38
  • Replacing one, two, and three pool models yields mean direction-corrected MM changes of −1.19%, −4.65%, and −4.27%; composition effects vary by task.Pool cardinality alone does not determine performance.Source: §D.1 · p. 38, Table 27 · p. 39

D.2 Modality-Utility Results

The section quantifies modality contributions and pair interactions using a fixed model’s predictive utility across modality subsets. Leading modalities vary by task, and individual contributions can differ from interaction effects.

D.2 Modality-Utility Results

  • Exact Shapley values quantify each modality’s contribution across all modality subsets of a fixed LLUNA model.Utility is negative absolute error on SIMS/MOSI and true-class or valid-answer log probability on other tasks; the empty-input reference predicts zero or a uniform distribution.Source: §D.2 · p. 40
  • Pair interactions measure departures from additive predictive utility, averaging the utility second difference over subsets of the other modalities.The reported medians and contribution frequencies describe variation across test examples, not training-seed uncertainty.Source: §D.2 · p. 40
  • The leading modality is audio on SIMS, text on MOSI, MELD, and URF, vision on Daily, TV auscultation on CirCor, and clinical inputs on ADNI.On WS, audio and vision contribute similarly; ADNI’s clinical and PET inputs contribute positively but have a −0.174-nat interaction.Source: §D.2 · p. 40
  • On WS, audio and vision contribute similarly; subtitles have a mean contribution of −0.007 nats, yet show positive interactions with both audio and vision, about +0.029 each.Source: §D.2 · p. 40

D.3 Chain-Sampling Results

The chain-sampling results compare sparse PCC training with full-lattice training and examine how sampled-chain count relates to performance. Sparse configurations retain comparable average missing-input performance with fewer comparisons, but more chains do not uniformly improve MM.

D.3 Chain-Sampling Results

  • Increasing sampled chains from one to three changes MELD MM from 54.35% to 53.41% and 54.44%, and URF MM from 69.55% to 69.13% and 69.08%.Thus, more sampled chains do not uniformly improve MM.Source: Figure 11 · p. 43, §D.3 · p. 43
  • Sparse configurations have an equal-task mean relative MM change of −0.28% versus full-lattice training, ranging from +1.40% on SIMS to −2.85% on CirCor.They use fewer PCC comparisons per training step, while testing still covers all nonempty availability states.Source: §D.3 · p. 43, Table 32 · p. 44
  • Sparse training reduces PCC comparisons from nine to two on MELD and from twenty-eight to three on CirCor; these counts measure training comparisons, not wall-clock time.Shared pairs are counted once, and test evaluation still includes every nonempty state.Source: §D.3 · pp. 43–44

E Motivation Diagnostics

  • The motivation diagnostics cover collaboration gains under modality removal, modality-specific feature recoverability, and changes in collaborative representations.These protocols underlie Figure 1 and the fixed-anchor analysis.Source: E Motivation Diagnostics · p. 44

E.1 Fixed-anchor collaboration gains

Fixed-anchor diagnostics track collaboration gains as inputs disappear, then examine modality-specific information recovery and state-dependent representations and predictions. LLUNA retains positive minimal-input mean gains, while the diagnostics document changes across availability states.

E.1 Fixed-anchor collaboration gains

  • Across eight tasks, the seven-baseline mean gain declines from Full to Minimal on all tasks and is negative at Minimal on six; LLUNA stays positive and exceeds it at both endpoints.The diagnostic uses the same fixed anchors, examples, and legal availability states for all seven baselines.Source: Table 33 · p. 45, §E.1 · p. 45
  • At Minimal, LLUNA’s four-method comparison has positive mean gains on all eight datasets, while the conventional subgroup’s mean is negative on all eight.These are dataset-level averages; individual anchor/view comparisons can remain negative.Source: §E.1 · p. 45, Table 34 · p. 46
  • After missing-input training, I2MoE, CLARGA, and Set Transformer have a mean gain that declines from Full to Minimal on every task and is negative at Minimal on URF, Daily, WS, and CirCor.This does not imply monotonic behavior for every intermediate transition or individual method.Source: §E.1 · p. 46, Table 35 · p. 47
  • Equally averaging domain-level absolute declines gives 9.32 pp for conventional collaboration and 5.79 pp for LLUNA, a 37.9% reduction.Medical-prediction declines are nearly unchanged: 6.51 versus 6.47 pp.Source: Table 40 · p. 49, §E.1 · p. 49

E.2 Private-signal recoverability

  • In the paired Attn. pool diagnostic, mean clipped-at-zero R2 recovery of the residual private-signal target was 58.52% for modality-alone representations versus 33.47% after full collaboration, decreasing on every dataset.The score measures linear recoverability of the residual, not task accuracy.Source: §E.2 · p. 50
  • All four evaluated collaboration methods have lower Full recovery under both training policies across 192 modality-level comparisons.Logit stack starts from lower recovery because it operates on modality logits; its smaller absolute decrease does not indicate better feature preservation.Source: §E.2 · p. 51, Table 41 · p. 51

E.3 Representation changes under modality removal

  • Across six three-modality tasks, mean representation distance is 0.279 for adjacent single-removal states and 0.456 for Full-to-Minimal states.Adjacent distances are smaller on every task with a multi-removal comparison; categories coincide on Daily.Source: §E.3 · p. 51

E.4 Prediction changes across availability states

  • Full-to-Minimal predictive JSD exceeds Adjacent JSD on all five classification tasks with a multi-removal contrast; Daily’s pair sets coincide.The diagnostic covers six classification tasks; SIMS and MOSI have scalar regression outputs.Source: §E.4 · p. 52, Figure 13 · p. 53

F Implementation and Reproducibility

The implementation specifies how model compatibility is estimated and used to construct the collaboration graph, alongside the prediction interfaces and PCC configuration. Reproducibility details include dataset-specific settings and released artifacts.

F Implementation and Reproducibility

  • The released code is intended to include split files, preprocessing, encoders, availability-state definitions, search spaces, seeds, and commands to reproduce every table.Optimization, regularization, early stopping, and state sampling follow per-dataset configurations.Source: F Implementation and Reproducibility · p. 53
  • Compatibility weights are estimated from training-partition predictions, then fixed during collaboration training and inference.Classification uses one minus half the L1 distance between probability vectors; regression uses the exponential of the negative absolute standardized prediction gap.Source: F Implementation and Reproducibility · p. 53
  • For each availability state, the graph is restricted to available models and incoming weights are normalized with a numerical stabilizer.Optional temperature scaling transforms agreement weights before union top-k pruning; positive self-loops are retained.Source: F Implementation and Reproducibility · p. 53

F.1 PCC weighting and discrepancy configuration

  • PCC compares normalized common consensuses of the same example across two availability states using cosine distance.Its base distance can be shaped by a dead zone and power, while edge weights normalize contributions within child missingness levels.Source: §F.1 · p. 54
  • Prediction heads use an explicit availability mask, while the private path retains direct access to adapted modality features.For each PCC edge, the richer endpoint is detached and the poorer endpoint receives the direct alignment gradient.Source: §F.1 · p. 54
  • Group readout uses a learned scoring vector and normalized attention weights over available models.Both sums run only over available models.Source: §F.1 · p. 54

G Additional Derivations

The derivations connect single-modality paths between availability states to bounds on common-representation and complete-prediction drift. The bounds depend on local cosine discrepancies, path length, and stated nonzero-consensus and Lipschitz assumptions.

G.1 Paths between arbitrary modality subsets

  • Single-modality transitions form a shortest path between any two distinct nonempty modality subsets, including equal-size states with different modalities.Each modality in the symmetric difference must change availability at least once, and the constructed path keeps every intermediate state nonempty.Source: §G.1 · p. 54

G.2 Local-to-global representation stability

  • The local-to-global bound controls endpoint common-representation distance using path length and accumulated adjacent cosine discrepancies, when every consensus is nonzero.Taking the minimum over admissible paths gives the tightest pathwise bound; longer paths with smaller discrepancies can provide tighter control.Source: §G.2 · p. 55
  • Weighted PCC objectives control the pathwise base distances when included edges have positive weights, with expected weights providing the corresponding control under random chain sampling.The stated stability result uses exact unit normalization; the epsilon-stabilized implementation is covered when path consensus norms meet the normalization floor.Source: §G.2 · p. 55

G.3 Supplementary output-drift bound

  • The complete-prediction drift bound separates changes in direct evidence from changes in the raw common consensus, under block-Lipschitz conditions.The private-path term includes availability conditioning, while pathwise PCC distances enter through the common representation's directional change.Source: §G.3 · p. 56
  • The common-consensus contribution to prediction drift depends on both magnitude and direction changes, with the pathwise stability result bounding the directional component.The derivation inserts an intermediate prediction that holds the common consensus fixed while changing the direct-evidence input.Source: §G.3 · p. 56

H Limitations

The evaluation is limited to controlled removal of whole modalities with known availability states, and the collaboration module is trained for a specified pool of frozen models. Broader-impact discussion notes that improved missing-modality behavior does not itself establish calibration, fairness, or safety.

H Limitations

  • Evaluation covers controlled removal of entire modalities with known availability states, not partial corruption, temporal misalignment, or sample-associated natural missingness.The controlled setting isolates effects across modality combinations.Source: H Limitations · p. 56
  • LLUNA is trained for a specified pool of frozen models; tested alternative pools require training the corresponding collaboration modules.Supporting pool changes without retraining the full collaboration module remains future work.Source: H Limitations · p. 57

I Broader Impact

  • Improved missing-modality behavior alone does not ensure calibration, fairness, or safety, which should be audited separately.The passage also recommends exposing the available-modality state to downstream users.Source: I Broader Impact · p. 57

Related papers

Loading 2610.20526v1…