PaperlayerView exact revision on arXiv

Source-linked AI summary

The Alignment Illusion in Multimodal Large Language Models

Hong-Han Wang, Yuntao Wang, Hu Ding

arXiv:2609.30210v1cs.CVcs.LG

TL;DR

Internal visual-text similarity in MLLMs is often read as evidence of content-level cross-modal interaction, though that interpretation has not been directly tested. The paper intervenes on visual inputs and analyzes shared-pathway geometry. It finds that scalar scores can remain high despite accuracy collapse, while the PA gap better tracks graded corruption and exposes divergence under irrelevant images.

  • Problem

    Whether internal alignment scores in MLLMs indicate content-level cross-modal interaction remains untested, despite their use in model design, data curation, and training diagnostics.

    Source: §1 · p. 1

  • Method

    The paper tests alignment with controlled visual-token interventions and introduces the PA gap to distinguish one-directional similarity from multi-directional visual structure.

    Source: §1 · p. 2

  • Results

    Across 13 MLLMs, removing visual content sharply reduces accuracy while standard scalar scores remain high, and the PA gap tracks graded corruption more consistently.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Internal alignment is a geometric diagnostic of the visual stream, not direct evidence that it supports the paired question, and is most informative alongside task accuracy under controlled interventions.

    Source: §4.3 · p. 9, §5 · p. 10

    The analysis covers projector-LLM architectures, controlled multiple-choice evaluation, and projector-output interventions; other fusion designs and task settings remain outside its scope.

    Source: §5 · p. 10

Abstract

from arXiv · show

Layer-wise visual-text similarity in Multimodal Large Language Models (MLLMs) is widely interpreted as evidence that the language model progressively integrates visual content into a shared representation space. This reading rests on the assumption that scalar alignment scores reflect content-level cross-modal interaction. To test this assumption, we apply controlled interventions to the visual stream. Across 13 MLLMs from five families spanning 0.5B to 72B parameters, replacing projector-output visual tokens with Gaussian noise sharply reduces task accuracy, yet four standard scalar measures (CKA, SVCCA, MIR, and the leading principal-angle cosine) fail to consistently separate the corrupted stream from the original. We call this failure the alignment illusion and trace it to the shared language-model pathway: anisotropic MLP down-projections pull visual and text tokens toward common output directions, producing weight-induced alignment. Because this component is essentially one-dimensional, we introduce the principal-angle gap (PA gap), defined as the difference between the top two principal-angle cosines, which separates weight-induced similarity from multi-directional visual structure. Under graded visual corruption, the PA gap tracks task accuracy more consistently than the scalar scores we consider; under a structured but irrelevant image, it further exposes regimes in which internal geometry and task accuracy come apart. Internal visual-text alignment in MLLMs is therefore best read as a geometric diagnostic of the visual stream inside the language model rather than a direct proxy for content-level cross-modal interaction, and is most informative when calibrated by controlled task evidence.

1 Introduction

The paper tests whether internal visual-text similarity in MLLMs demonstrates content-level cross-modal interaction. Controlled visual interventions reveal that standard scores can remain high without reliable use of visual content, motivating a geometric measure calibrated against task accuracy.

1 Introduction

  • The conventional interpretation that rising layer-wise similarity signals progressive visual integration relies on an untested assumption about content-level interaction.That interpretation has informed model design, data curation, and training diagnostics.Source: §1 · p. 1
  • Shared MLLM weights can impose similar visual and text geometry without task-relevant visual content, which scalar scores conflate with content-level interaction.The paper calls this failure mode the alignment illusion.Source: §1 · p. 2
  • Across 13 MLLMs, removing visual content sharply reduces task accuracy while standard scalar alignment scores remain high.The study tests Gaussian-noise replacement and structured irrelevant images.Source: §1 · p. 2
  • The principal-angle gap separates one-directional weight-induced similarity from multi-directional visual structure and tracks accuracy more consistently under graded corruption.With structured irrelevant input, it also exposes cases where internal geometry and task accuracy diverge.Source: §1 · p. 2
  • Prior work examines representation similarity, visual dependence, anisotropy, and multimodal mechanisms, but does not test when MLLM alignment supports a content-interaction interpretation.This paper frames its contribution as addressing that validity question.Source: §1 · pp. 2–3

2 Shared Processing and Alignment Geometry

The paper defines alignment on hidden states produced by a shared projector-LLM pathway, then compares visual and text subspaces layer by layer. Its principal-angle spectrum provides both a leading scalar score and the fuller geometry used in later analyses.

2.1 Projector-LLM Architecture

  • Projector-LLM models map vision features into language-model embeddings, concatenate visual and text tokens, and process both through shared Transformer blocks.The resulting shared-pathway hidden states are the inputs to all alignment scores in the paper.Source: §2.1 · p. 3, §2 · p. 3

2.2 Principal-Angle Spectrum and Standard Alignment Scores

  • The principal-angle-cosine spectrum compares top-k PCA subspaces of visual and text tokens; its leading cosine is a scalar score, while the full spectrum captures remaining angles.The paper uses k = 30 unless stated otherwise.Source: §2.2 · p. 3
  • Figure 2 presents ORIG-to-NOISE accuracy drops, layer-wise leading principal-angle cosines, and scalar-score separation across models.Its separation score is positive when a measure ranks ORIG above NOISE.Source: Figure 2 · p. 4

3 The Alignment Illusion

Controlled visual-token interventions show that standard scalar alignment scores can remain high after visual content is removed, while task accuracy falls. The paper traces this mismatch to shared language-model weights and introduces the PA gap to distinguish one-directional weight-induced alignment from richer visual structure.

3.1 Noise Leaves Alignment Scores Nearly Unchanged

  • Under NOISE, task accuracy drops by 38 to 50 pp across all 13 models, while the leading principal-angle cosine changes little and can approach unity.NOISE replaces visual tokens with norm-matched Gaussian vectors, preserving scale while removing visual content.Source: §3.1 · p. 4
  • Across 13 models, σ1, CKA, and SVCCA have negative median separation scores for ORIG versus NOISE, while MIR is near zero.The four standard scalar measures therefore provide little or no reliable separation between original and content-removed visual streams.Source: §3.1 · p. 4

3.2 Shared Language-Model Weights Create Apparent Alignment

  • Bypassing the MLP changes σ1 3.5× more than bypassing attention, and ORIG and NOISE MLP-output subspaces share a median leading cosine of 0.996.This supports an MLP contribution to apparent alignment even when the visual stream is replaced by noise.Source: §3.2 · p. 5
  • Across 13 models, s1(Wout)/s2(Wout) is 1.2 to 1.6 and exceeds the matched random reference; the principal-angle basis captures 2.8× to 13.6× more energy than random at r=10.The results locate principal-angle directions within dominant Wout output directions.Source: §3.2 · p. 5, Figure 3 · p. 5
  • A shared down-projection can inflate downstream alignment because it pulls both input outputs toward a common direction, without coupling the inputs’ content.The bound’s separate terms depend on each input and the shared weight, not on a content-level interaction between the inputs.Source: §3.2 · p. 6

3.3 The Principal-Angle Gap

  • The PA gap is the difference between the top two principal-angle cosines: small gaps indicate multi-directional alignment, while large gaps indicate alignment mainly along Wout’s direction.This targets the distinction between structured visual content and a single architecture-induced direction.Source: §3.3 · p. 6
  • ∆PA is larger under NOISE across all models and is the only measure that ranks ORIG above NOISE in all 13 models.By contrast, the layer-averaged σ1 difference is close to zero and can reverse direction.Source: §3.3 · p. 6

3.4 The PA Gap Tracks Graded Loss of Visual Content

  • Under α-mixing, the PA gap has the strongest reported accuracy correlation: mean |r| = 0.894, median |r| = 0.917, and minimum |r| = 0.655 across 13 models.Its |r| exceeds 0.80 in 12/13 models and has the expected sign in all 13; the comparison uses the inner 80% of model layers.Source: §3.4 · p. 7, Table 1 · p. 7

4 Structured but Irrelevant Visual Content

The section tests how structured but irrelevant visual input affects accuracy and internal geometry, then compares both with the effects of removing visual structure. The PA gap distinguishes the geometry of original, irrelevant, and noise streams, but task accuracy does not follow that ordering.

4.1 Irrelevant Images Are Actively Processed

  • Across 13 models, IRR accuracy is a median 5.0 pp below TEXT, while NOISE is 2.1 pp below TEXT; shuffled tokens differ from ORIG by only 0.8 pp.The IRR-versus-TEXT gap is 2.9 pp for models under 3B parameters.Source: §4.1 · pp. 7–8, Table 2 · p. 8

4.2 Localizing the Irrelevant-Image Effect

  • Projector probes recover the irrelevant-image category at 74.4–78.5% versus 5% chance, and matched in-band perturbations hurt more in all 13 models by 1.1–5.5 pp.At ε=1.0, the in-band versus out-of-band accuracy-drop difference isolates the effect attributable to principal-angle directions.Source: Figure 5 · p. 8, §4.2 · p. 8

4.3 Geometric Ordering and Accuracy Divergence

  • PA-gap medians rise from 0.130 (ORIG) to 0.213 (IRR) to 0.324 (NOISE), ordering the three streams by increasingly one-directional geometry.The same full ordering holds in almost all models; standard scalar measures do not order the settings consistently.Source: §4.3 · pp. 8–9, Table 3 · p. 9
  • Median accuracy is 85.4% under ORIG, versus 36.1% under IRR and 39.0% under NOISE, so geometry and accuracy order the streams differently.IRR is geometrically more structured than NOISE, but is slightly less accurate for the paired question.Source: §4.3 · p. 9

5 Discussion

Internal visual-text alignment is a geometric diagnostic, not direct evidence that visual content supports the paired question. The discussion recommends interpreting alignment alongside task accuracy under controlled interventions and notes the study’s scope boundaries.

5 Discussion

  • Alignment scores can conflate content-dependent structure with weight-induced similarity, so the geometry producing a score matters as much as its magnitude.The PA gap separates these sources and orders ORIG, IRR, and NOISE into distinct geometric regimes.Source: §5 · p. 9
  • Alignment scores are most reliable alongside task-accuracy evidence under controlled visual interventions.Internal geometry alone does not establish whether the visual stream supports the paired question.Source: §5 · pp. 9–10
  • The analysis covers projector-LLM architectures, controlled multiple-choice evaluation, and projector-output interventions, not other fusion designs or task types.The paper identifies open-ended generation, video, document, and agentic tasks as directions for future analysis.Source: §5 · p. 10

A Evaluation protocol and statistical tests

The appendix documents the shared evaluation protocol and defines how scalar scores are compared between ORIG and NOISE. Accuracy uses the same questions across settings, with norm checks and statistical procedures supporting the comparisons.

A Evaluation protocol and statistical tests

  • Task accuracy uses the same 1,000 MMBench questions across all models and interventions, while five settings vary the visual stream from unmodified tokens to noise, irrelevant images, shuffled tokens, or text only.Because the questions are fixed, accuracy differences across settings reflect the intervention rather than a change in evaluation data.Source: A Evaluation protocol and statistical… · p. 12
  • The per-sample IRR/ORIG projector-output norm ratio has median in [0.994, 1.002] across 13 models, supporting a check that intervention effects are not simply signal-magnitude differences.NOISE, IRR, and SHUF preserve or closely match ORIG per-token norms.Source: A Evaluation protocol and statistical… · p. 13, Table 5 · p. 13
  • Accuracy uncertainty is estimated with 1,000 bootstrap resamples, and paired comparisons against ORIG use McNemar’s test with the corrected-χ2 or exact form according to discordant-pair count.SHUF has point-estimate accuracy only and no paired test because its run did not store per-sample predictions.Source: A Evaluation protocol and statistical… · p. 13, Table 6 · p. 13, Table 7 · p. 14
  • Per-model behavioral values underlie the compact median/IQR contrasts, with setting-minus-reference differences coded so accuracy drops are negative.Table 8 reproduces the five contrasts used in the main table.Source: A Evaluation protocol and statistical… · p. 13, Table 8 · p. 14

B Summary separation score

  • The scalar separation score min–max normalizes each summary across settings and layers, averages the ORIG–NOISE separation over the inner 80% of layers, and treats scores near zero or below as little or no separation.The first and last 10% of layers are excluded because the embedding layer is dominated by input geometry and the final layer by the output head.Source: B Summary separation score · pp. 14–15

C Model specifications

The evaluation covers 13 released-weight MLLMs spanning five families, without fine-tuning. Across models, the leading principal-angle cosine stays high under noise, while the PA gap is larger under NOISE.

C Model specifications

  • The 13 evaluated MLLMs span five families, use released weights without fine-tuning, and vary only in the projector-output intervention across settings.The families are LLaVA-OV, LLaVA-OV-1.5, Qwen2-VL, Qwen2.5-VL, and InternVL3.Source: C Model specifications · p. 15, Table 10 · p. 15

D Per-model σ1 and ∆trajectories under ORIG and NOISE

E Full α-interpolation results

Under α-interpolation from NOISE to ORIG, the PA gap tracks graded visual degradation across models more consistently than the scalar scores considered. The appendix reports its per-model correlations and displays accuracy and geometry across the interpolation sweep.

E Full α-interpolation results

  • The α-interpolation plots pair task accuracy with PA gap across a grid where α = 0 is NOISE and α = 1 is ORIG.The accuracy and PA-gap series use separate axes, and subplots are grouped by model family.Source: Figure 8 · p. 17
  • The comparison includes PA gap, σ1, participation ratio, spectral entropy, CKA, SVCCA, and MIR, with per-model correlations summarized in Table 12.Figure 9 shows the inner-80%-layer-trimmed mean |r| for each model and scalar summary.Source: Figure 9 · p. 18, Table 12 · p. 18
  • Mean |r| = 0.894, median |r| = 0.917, and minimum |r| = 0.655 for the PA gap’s association with task accuracy; |r| exceeds 0.80 in 12/13 models.These are inner-80%-layer-trimmed mean absolute correlations under α-interpolation.Source: E Full α-interpolation results · p. 16, Table 12 · p. 18

F Mechanistic controls

Mechanistic controls implicate shared MLP output geometry in the similarity observed under NOISE. Bypass comparisons, weight-spectrum measurements, sensitivity tests, and random initialization further characterize this mechanism and its scope.

F Mechanistic controls

  • 0.036 (MLP) versus 0.010 (attention) is the cross-model median downstream |∆σ1| effect, a 3.5× larger effect for MLP bypass across all 13 models.The per-model comparison uses the same target layers for both bypasses.Source: F Mechanistic controls · p. 19, Figure 10 · p. 19, Table 13 · p. 19
  • Across 12 models with available spectra, the layer-averaged s1(Wout)/s2(Wout) ratio ranges from 1.166 to 1.635, with median 1.355 and IQR 1.329–1.439; all 13 values exceed the matched random reference.The passage reports 12 available spectra but states that all 13 values exceed the reference.Source: F Mechanistic controls · p. 19
  • The principal-angle basis captures 2.770–13.574 times the matched-random-baseline energy in Wout’s top-10 left singular subspace, with median ratio 9.972 across 13 models.All 13 ratios exceed 1.Source: F Mechanistic controls · p. 20
  • The σ1 trajectory shape is stable across PCA dimensions k ∈ {5, 10, 15, 20, 30, 50} in the two tested models, with k = 30 as the default.The two models are OV-Q2-0.5B and Q2.5-VL-3B.Source: F Mechanistic controls · p. 20, Figure 12 · p. 21
  • With random initialization, mid-layer mean σ1 under NOISE drops from 0.745–0.987 for trained weights to 0.615–0.736, while ORIG accuracy is 11.6–24.3%.The control covers four models; the authors interpret the result as showing that inflated σ1 under NOISE arises from trained weight structure rather than a generic architectural property.Source: F Mechanistic controls · p. 20, Table 15 · p. 21

G Additional results for the principal-angle / Wout overlap

The principal-angle basis captures substantially more energy than a random baseline in Wout’s top singular subspace, and the appendix reports how this overlap changes across subspace sizes.

G Additional results for the principal-angle / Wout overlap

  • 2.8× to 13.6× more energy than a random baseline is captured by the principal-angle basis in Wout’s top-10 left singular subspace.The appendix extends this comparison across r, from small subspaces to the full available Wout output subspace.Source: G Additional results for the… · p. 21
  • The expected full-sweep pattern is that principal-angle energy exceeds the matched random baseline at small r, then decays toward 1× as r grows.The figure’s solid lines denote the principal-angle basis; dashed lines denote matched-dimension random baselines.Source: G Additional results for the… · p. 21, Figure 13 · p. 22
  • Figure 13 compares projected energy across r using principal-angle-basis curves and matched-dimension random-baseline curves.The caption identifies the line styles but does not report additional numerical sweep values.Source: Figure 13 · p. 22

H Noise injection protocol and causal sweep

The appendix details controlled residual-stream noise injections and image-category probes, reporting stronger in-band than out-of-band effects at ε = 1.0 and retained information about irrelevant images. It also specifies the rule for ranking scalar metrics across settings.

H Noise injection protocol and causal sweep

  • Noise is injected into visual-token residual activations along or orthogonal to the principal-angle basis, scaled relative to each activation; text tokens remain unperturbed.In-band noise uses normalized random combinations of basis directions, while out-of-band noise is projected orthogonally to that span.Source: H Noise injection protocol and causal… · pp. 22–23
  • 1.12–5.48 percentage points, with median 2.38 and IQR 1.86–3.11, is the in-minus-out accuracy-drop difference at ε = 1.0 across 13 models.In-band perturbations induce larger drops than out-of-band perturbations in all 13 models; at ε = 0.3 the direction persists but the maximum difference is under one pp.Source: H Noise injection protocol and causal… · p. 23, Table 16 · p. 23, Figure 14 · p. 23

I Linear probe protocol

  • 74.4%–78.5% own-category accuracy (median 76.4%) versus 5.5%–7.2% question-category accuracy (median 6.4%) shows that projector outputs retain the irrelevant image’s category, not the answer-image category.The probes use 1,000 pairs with 20 balanced categories; chance is 5%.Source: I Linear probe protocol · p. 24, Table 17 · p. 24

J Per-setting scalar summaries and ranking consistency

  • A metric ranks ORIG above a comparison setting only when its difference reaches 5% of the per-model range; larger values favor alignment for σ1, CKA, and SVCCA, while lower values do so for MIR and the PA gap.Table 18’s per-model scalar summaries are averaged over the inner 80% of layers and used to derive Table 3.Source: J Per-setting scalar summaries and… · p. 24, Table 18 · p. 25

K Proof of Proposition 1

The proof formalizes how shared output directions can raise leading visual-text alignment without content-level coupling. It bounds the score using the angles between each output’s leading direction and a common direction.

K Proof of Proposition 1

  • The leading principal-angle cosine is evaluated between output subspaces, with their leading output directions serving as feasible unit vectors in those subspaces.This feasibility supplies the starting comparison in the proof’s variational argument.Source: K Proof of Proposition 1 · p. 25
  • If both leading output directions lie within angles θX and θY of a shared direction U1, the triangle inequality bounds their mutual angle; the proof handles the cases where θX + θY is at most or greater than π/2.When the sum exceeds π/2, nonnegativity of the leading cosine makes the bound trivial.Source: K Proof of Proposition 1 · pp. 25–26
  • A shared down-projection with dominant output directions can increase leading visual-text alignment without task-relevant content coupling between the inputs.This is the paper’s formal interpretation of weight-induced alignment.Source: K Proof of Proposition 1 · p. 26
Loading 2609.30210v1…