The Alignment Illusion in Multimodal Large Language Models
Hong-Han Wang, Yuntao Wang, Hu Ding
cs.CV · cs.LG
Quick verdict
Argues MLLM visual-text alignment scores remain high under visual corruption via weight-induced anisotropy; PA gap is convincing but lacks whitening baselines.
Summary
The paper asks whether layer-wise visual-text similarity in multimodal LLMs actually evidences content-level cross-modal integration. Across 13 projector-LLM models from five families (0.5B–72B), the authors replace projector-output visual tokens with norm-matched Gaussian noise: accuracy on a 1,000-question MMBench subset drops 38–50 points, yet CKA, SVCCA, MIR, and the leading principal-angle cosine fail to separate the corrupted stream from the original — the 'alignment illusion.' They trace the effect to the shared language-model pathway in four steps: alignment amplifies inside the LLM, localizes to the MLP (bypass effects 3.5× attention), MLP down-projections show anisotropy (s1/s2 ≈ 1.2–1.6, above random), and the principal-angle basis overlaps the down-projection's top singular subspace (2.8–13.6× energy); Proposition 1 formalizes that σ1 can inflate without input coupling. The proposed principal-angle gap (σ1−σ2) ranks ORIG > IRR > NOISE in 13/13 models and tracks accuracy under graded α-mixing (mean |r|=0.894, 12/13 above 0.80), beating CKA, SVCCA, MIR, participation ratio, and spectral entropy. Irrelevant-image experiments show structured-but-useless content is actively processed (in-band perturbations hurt more in 13/13) while lowering accuracy below the text-only baseline. Limitations: single multiple-choice benchmark, projector architectures only, one shuffled-token run lacks per-sample predictions, no code release stated, and anisotropy-corrected (whitened) similarity is never tested as a baseline.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Standard scalar alignment measures (CKA, SVCCA, MIR, leading principal-angle cosine) fail to separate original from content-removed visual streams across 13 MLLMs — the alignment illusion.
- The illusion is caused by weight-induced alignment: anisotropic MLP down-projections impose common one-directional output geometry on both modalities (Figure 3 steps 1–4, Proposition 1).
- The proposed principal-angle gap (σ1−σ2) separates weight-induced from multi-directional alignment, ordering ORIG > IRR > NOISE in 12–13/13 models and tracking graded corruption best (mean |r|=0.894, 13/13 sign-consistent).
- Structured but irrelevant images are actively processed through the same principal-angle subspace and can reduce accuracy below the text-only baseline, showing internal geometry and task usefulness come apart.
Strengths
Causally grounded validity test with tight controls
Five visual-stream settings (ORIG/NOISE/IRR/SHUF/TEXT, Appendix A) hold per-token norms constant (Table 5, IRR/ORIG ratios 0.994–1.002), and a random-initialization control (Table 15) shows inflated σ1 under NOISE depends on trained weights, not architecture.
Broad, fully specified model coverage
13 models from five families spanning 0.5B–72B (Table 10), same fixed 1,000-question MMBench subset and seed 42 across all settings; all 13 models show the 38–50pp accuracy collapse under NOISE (Figure 2A).
Multi-step mechanistic evidence converging on one cause
Strong behavioral statistics
Honest scoping and self-limiting interpretation
Weaknesses
Missing anisotropy-corrected similarity baseline
Single benchmark, single task format
SHUF condition lacks per-sample statistics
No significance test on the key metric comparison
PA gap lacks absolute calibration
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets a specific validity question: does layer-wise visual-text representational similarity in projector-based MLLMs actually evidence content-level cross-modal integration (Section 1)? Across 13 models from five families (LLaVA-OV, LLaVA-OV-1.5, Qwen2-VL, Qwen2.5-VL, InternVL3; 0.5B–72B, Table 10), the authors test this with controlled visual-stream interventions and conclude: (1) four standard scalar measures (σ1, CKA, SVCCA, MIR) fail to separate original from content-removed streams — Figure 2C separation-score medians are negative for σ1/CKA/SVCCA and +0.013 for MIR; (2) the cause is weight-induced alignment via anisotropic MLP down-projections; (3) the proposed PA gap (σ1−σ2, Definition 2) is a better diagnostic. Claims are appropriately scoped: the Discussion restricts conclusions to projector-LLM architectures and multiple-choice evaluation.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the 'alignment illusion' in multimodal LLMs?
It is the finding that standard internal visual-text similarity scores (CKA, SVCCA, MIR, and the leading principal-angle cosine) stay high even when the model's visual input is replaced by content-free noise and task accuracy collapses by tens of points. High internal alignment can therefore be produced by the network's shared weights rather than genuine use of visual content.
What is the principal-angle gap (PA gap)?
A diagnostic defined as the difference between the two largest principal-angle cosines (σ1−σ2) between visual and text token subspaces at each layer. A large gap means alignment is concentrated in one weight-imposed direction; a small gap means alignment spans multiple directions, which corresponds to structured visual content. In the paper it ordered original > irrelevant > noise images consistently in nearly all 13 tested models.
Which models were tested in this paper?
13 open-source multimodal LLMs from five families — LLaVA-OneVision, LLaVA-OneVision-1.5, Qwen2-VL, Qwen2.5-VL and InternVL3 — spanning 0.5B to 72B parameters, all projector-based, evaluated on a fixed 1,000-question subset of MMBench.
Does high visual-text alignment mean the model actually uses the image?
Not by itself, according to this paper. Internal alignment is best read as a geometric diagnostic of the visual stream inside the language model; claims that a model uses visual content are most reliable when paired with task-accuracy evidence under controlled visual interventions.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–14 are described via captions and surrounding text but their pixel content cannot be verified from the extracted text.
- No code/repository link appears in the extracted text; availability was judged from the paper body only.
- All appendices A–K and tables 4–18 appear fully extracted, so extraction-related uncertainty is low.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.