PaperlayerView exact revision on arXiv

Source-linked AI summary

Architecture-Dependent Fusion Pathways in MLLMs

Hebao Zhu, Dongxia Wu

arXiv:2610.03289v1cs.LG

TL;DR

Although MLLMs perform strongly on vision-language tasks, how visual and textual information fuses across layers remains insufficiently understood, particularly across architectural paradigms. The paper analyzes representative concatenation and native models through complementary representation, attention, and dimensionality analyses, with separate causal interventions. It reports text-first, vision-later fusion in concatenation models and earlier visual-textual co-adaptation in native models.

  • Problem

    Downstream task performance does not reveal how visual and textual tokens interact across depth, and fusion dynamics have not been systematically characterized across MLLM architectures.

    Source: §1 · p. 1

  • Method

    The study analyzes representative architectures through alignment decoupling, attention routing and entropy, and intrinsic dimensionality, with separate causal interventions to assess functional relevance.

    Source: §1 · p. 2, §3.1 · p. 4, §3.2 · p. 4

  • Results

    Concatenation models follow a text-first, vision-later pathway, whereas native models show earlier visual-textual co-adaptation, earlier feature-space expansion, and heterogeneous attention profiles.

    Source: §3.6 · p. 8

  • Takeaways & Limitations

    The observed pathway differences can guide transition-layer and sparse-anchor robustness diagnostics and architecture-aware placement of layer-selective fine-tuning or adapters.

    Source: §5 · p. 9

    Cross-modal CKA alone cannot identify which modality changes, how information is routed, or whether representation-level transitions are functionally important; a perfectly controlled comparison across architectures is not feasible.

    Source: §1 · p. 2, §3.1 · p. 4

Abstract

from arXiv · show

Multimodal Large Language Models (MLLMs) achieve strong performance across vision-language tasks, yet the internal mechanisms by which visual and textual information are fused across layers remain insufficiently understood. We investigate representative MLLMs from two architectural paradigms: concatenation architectures and native multimodal architectures. We conduct three progressively connected analyses: alignment decoupling identifies which modality changes, attention routing and entropy characterize how cross-modal information is distributed, and intrinsic dimensionality examines how fusion reshapes feature spaces. Separately, we perform causal intervention experiments as a validation of the resulting interpretation. As a supplementary analysis, we use visual CKA to examine the Platonic Representation Hypothesis. Together, these analyses reveal two distinct fusion pathways: concatenation models follow a text-first, vision-later pathway, whereas native models exhibit earlier visual-textual co-adaptation and feature-space reorganization. This work provides a mechanistic perspective for understanding multimodal fusion and supports architecture-aware diagnostics of multimodal representations.

1 Introduction

Downstream performance does not show how visual and textual tokens interact across model depth, and fusion dynamics have not been systematically characterized across MLLM architectures. The study combines layer-wise representation, attention, and dimensionality analyses to describe distinct fusion pathways, with separate causal interventions validating their functional relevance.

1 Introduction

  • Downstream accuracy and reasoning benchmarks do not reveal how modalities fuse across depth, and layer-wise fusion dynamics remain uncharacterized across architectural paradigms.The paper distinguishes concatenation and native multimodal architectures by how they integrate visual and textual information [6].Source: §1 · p. 1
  • Concatenation models show relatively high early cross-modal similarity followed by a pronounced intermediate decline, while native models change earlier and generally recover similarity in middle-to-deep layers.Native models may show model-specific declines near their final layers.Source: §1 · pp. 1–2
  • Cross-modal CKA cannot identify which modality changes, how information is routed, whether feature spaces compress or reorganize, or whether transition layers are functionally important.The paper therefore treats CKA as insufficient for explaining fusion mechanisms on its own.Source: §1 · p. 2
  • The study links alignment decoupling, attention routing and entropy, and intrinsic dimensionality, then separately uses causal interventions to test functional relevance.A supplementary visual CKA analysis examines the Platonic Representation Hypothesis [12].Source: §1 · p. 2

2 PRELIMINARY

MLLM architectures differ in how they integrate visual and textual information, while their fusion dynamics across depth have not been comprehensively characterized. Prior interpretability work has examined information flow or alignment, but often focuses on concatenation architectures, isolated layers, or representation alignment alone.

2.1 MLLM Architectures

  • Concatenation architectures retain the language-model backbone and use projection modules to align visual features with textual embeddings, often relying on pretrained CLIP visual encoders [13].Their core assumption is that the visual encoder already supplies sufficiently high-level visual semantics [14].Source: §2.1 · p. 3
  • Native architectures learn visual tokens and language representations in a shared end-to-end framework, with visual front ends that may be jointly optimized or encoder-free.The category is defined by the multimodal learning interface rather than a single tokenization implementation.Source: §2.1 · p. 3

2.2 MLLM Alignment and Interpretability

  • Prior work has not comprehensively explained fusion across depth, with studies often focused on concatenation models, representation alignment, or isolated layer representations [19] [20] [9].The section also notes that stronger alignment does not always lead to better performance [18].Source: §2.2 · p. 3

3 Experiments

Across six representative models, concatenation architectures show a text-first, vision-later pathway, while native models show earlier visual-textual co-adaptation and feature-space reorganization. Separate causal interventions support the functional relevance of these pathway interpretations.

3.1 Experimental Setup

  • The study analyzes representative concatenation and native models using alignment, attention, and intrinsic-dimensionality measures, then validates the interpretation with interventions.The comparison includes three models per paradigm; the authors note that their different training pipelines preclude fixing every factor in a controlled comparison.Source: §3.1 · pp. 3–4

3.2 Alignment Decoupling

  • Concatenation models retain visual representations in shallow layers as textual representations adapt earlier, then reshape visual representations later.Native models diverge visually earlier and show model-specific textual evolution with depth; concatenation models also maintain higher visual CKA at layers 8 and 15.Source: §3.2 · pp. 4–5, Figure 3 · p. 5

3.3 Attention Routing and Entropy

  • Concatenation models have visual-to-text normalized entropies of 0.046–0.192, versus 0.497 for EVE and 0.301 for SOLO, indicating more concentrated attention in that direction.Visual-to-text routing exceeds reverse-direction routing across displayed models, but routing magnitude alone does not clearly distinguish the architectural families.Source: §3.3 · p. 6, Figure 4 · p. 6

3.4 Intrinsic Dimensionality Transition

  • Cambrian’s visual intrinsic dimensionality stays low and stable through early and transition layers before increasing, while EVE’s visual dimension begins increasing earlier.Cambrian’s textual dimension decreases from its input-side value and stabilizes; EVE’s textual dimension changes with the visual side and can peak before decreasing.Source: §3.4 · p. 7, Figure 5 · p. 7

3.5 Causal Intervention Validation

  • Native models are more vulnerable to early-layer noise, whereas concatenation models are more vulnerable to late-layer noise; masking concentrated textual targets causes larger POPE drops in concatenation models.The authors interpret these patterns as consistent with early native co-adaptation and later visual reshaping in concatenation models.Source: §3.5 · pp. 7–8, Table 1 · p. 8, Table 2 · p. 8

3.6 Mechanistic Summary

  • Concatenation models follow a text-first, vision-later pathway, while native models show earlier visual-textual co-adaptation and feature-space reorganization.The intervention results separately support the functional relevance of both pathway interpretations.Source: §3.6 · p. 8

4 Deep Visual Representation Convergence

The supplementary visual CKA analysis tests whether MLLMs with different architectures converge toward similar visual representations. It finds that shallow representations differ more across architectures, while deep representations are more similar.

4 Deep Visual Representation Convergence

  • The analysis computes pairwise CKA between models’ visual representation spaces at each layer, producing a model-by-model similarity matrix.It is motivated by the Platonic Representation Hypothesis and tests for convergence across models with different architectures.Source: §4 · p. 8
  • At layer 1, visual representations are architecture-dependent, while layer 30 shows substantially higher similarity across almost all model pairs.The authors interpret this as deep visual representations moving toward a more similar semantic structure.Source: §4 · p. 8, Figure 6 · p. 9
  • The results provide empirical evidence consistent with the Platonic Representation Hypothesis: deep visual representations are more aligned across models.From representation similarity, multimodal fusion appears to reduce architecture-specific differences in high-level visual representations.Source: §4 · pp. 8–9

5 Conclusion

The paper uncovers internal fusion mechanisms and proposes directions for diagnostics, training strategies, and architecture-aware fine-tuning.

5 Conclusion

  • The findings suggest using transition layers and sparse textual anchors as robustness diagnostics, and guiding training or layer-selective adaptation according to architecture-specific pathways.Proposed options include earlier cross-modal feedback for concatenation models and placing adapters or LoRA modules selectively across layers.Source: §5 · p. 9

A Technical Appendices and Supplementary Material

The appendices provide supporting analyses, fuller model-level results, and robustness checks for the main findings.

A Technical Appendices and Supplementary Material

  • Appendix A.1 verifies that the global CKA trend is reflected in local neighborhood structures, while A.2 provides individual-model visual and textual CKA curves.Source: A Technical Appendices and… · p. 12
  • Appendix A.3 reports complete attention-routing and cross-modal-entropy results, and A.4 adds intrinsic-dimensionality curves supporting feature-space expansion patterns.Source: A Technical Appendices and… · p. 12
  • Appendix A.5 shows that deep-layer visual representation convergence develops progressively across depth, and A.6 reports bootstrap robustness analysis.Source: A Technical Appendices and… · p. 12

A.1 Mutual KNN

The Mutual KNN analysis checks whether the global CKA trends also appear in local neighborhood structure. Its results show a clearer middle-layer transition in concatenation models and more continuous evolution in native models.

A.1 Mutual KNN

  • Mutual KNN consistency compares visual and textual nearest-neighbor structures at k = 10; higher consistency indicates more stable or better-aligned local semantic structure.Source: §A.1 · p. 12
  • The Mutual KNN results follow the global CKA trends: concatenation models show a more obvious middle-layer transition, while native models evolve more continuously.This supports the reliability of the global CKA observation in the main text.Source: §A.1 · p. 12, Figure 7 · p. 12

A.2 Per-model Intra-modality CKA

Figure 8 tracks each model’s visual and textual representations against their initial references. Visual trajectories differ by architecture, while textual changes vary across models.

A.2 Per-model Intra-modality CKA

  • Figure 8 separates concatenation and native models and plots visual and textual CKA relative to the initial reference across zero-based layers 0–32.Blue and orange curves denote visual and textual representations, respectively.Source: §A.2 · p. 12, Figure 8 · p. 13
  • Native models show stronger early visual changes than concatenation models, whose shallow-layer visual representations remain closer to the reference.The per-model curves are measured relative to the initial representation.Source: §A.2 · p. 13, Figure 8 · p. 13
  • Textual trajectories vary across models, including an early decrease followed by recovery in SOLO.The individual visual and textual curves do not imply synchronized or monotonic changes in both modalities.Source: §A.2 · p. 13

A.3 Attention Routing and Cross-modal Entropy

Layer-wise attention profiles characterize cross-modal routing and entropy over layers 8–15. Concatenation models and native models show different routing profiles, with heterogeneous patterns among native models.

A.3 Attention Routing and Cross-modal Entropy

  • The profiles cover layers 8–15 for all six models, including SAIL, whose constrained path was omitted from the main-text comparison.The main text compares distributions and directional median gaps for the other five models.Source: §A.3 · p. 13, Figure 9 · p. 14
  • Concatenation models combine relatively high visual-to-text routing with low entropy in that direction across the analyzed range.Figure 9 reports layer-wise routing ratios and per-sample normalized cross-modal entropy.Source: §A.3 · p. 13, Figure 9 · p. 14
  • Native models have heterogeneous attention profiles, including more diffuse visual-to-text attention in EVE and SOLO and a constrained entropy path in SAIL.SAIL has zero visual-to-text entropy because its hybrid attention path constrains that direction.Source: §A.3 · p. 13, Figure 9 · p. 14

A.4 Intrinsic Dimension

The intrinsic dimensionality analysis examines whether fusion corresponds to feature-space compression, expansion, or reorganization. Its supplementary curves show architecture-dependent timing of visual dimensionality expansion, consistent across variance-retention thresholds.

A.4 Intrinsic Dimension

  • The analysis uses per-model visual and textual curves and includes a threshold-sensitivity analysis for two representative models.Intrinsic dimensionality is used to characterize compression, expansion, or reorganization of feature spaces during fusion.Source: §A.4 · p. 14
  • Concatenation models generally show delayed visual dimensionality expansion, while native models show earlier and more continuous feature-space growth.The reported pattern is consistent with a text-first, vision-later pathway in concatenation models and earlier modality co-adaptation in native models.Source: §A.4 · p. 14
  • The observed dimensionality trends remain consistent across varying variance-retention thresholds, rather than depending on a specific variance cutoff.Source: §A.4 · p. 14

A.5 Cross-model Visual Representation Similarity

Cross-model visual representations are more architecture-dependent in shallow layers and more aligned in deeper layers. Bootstrap resampling finds the reported CKA patterns stable across sample sizes.

A.5 Cross-model Visual Representation Similarity

  • Pairwise visual CKA compares models at shallow and deep layers to examine whether deep-layer visual representations become more aligned.This supplementary analysis is motivated by the Platonic Representation Hypothesis.Source: §A.5 · p. 14
  • Deep-layer visual CKA is substantially higher across model pairs than shallow-layer similarity, where representations are strongly architecture-dependent.The heatmaps suggest that deep visual representations become more aligned after multimodal fusion.Source: §A.5 · p. 14, Figure 11 · p. 16

A.6 Bootstrap Robustness Analysis

  • The bootstrap uses image-text pairs as resampling units, four sample sizes from 500 to 5,000, and 1,000 resamplings per size.Relative bias is measured against the bootstrap mean for the 5,000-pair reference, not an unknown population CKA value.Source: §A.6 · p. 15
  • Bootstrap standard errors remain small across sample sizes, and relative bias decreases as sample size increases, supporting stability across shallow, transition, and deep layers.Figure 12 shows relative bias against the 5,000-sample estimate for representative layers.Source: §A.6 · p. 15, Figure 12 · p. 17
Loading 2610.03289v1…