Open Brain
arXivPaperNeeds Review

The Alignment Illusion in Multimodal Large Language Models

Common visual-text alignment scores stayed high after visual tokens were replaced with noise. Multimodal evaluations should pair internal geometry with controlled corruption and task accuracy.

arXiv · Sep 24, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Across **13 multimodal models** from five families, replacing visual tokens with Gaussian noise sharply reduced accuracy. Yet **four standard alignment measures** did not consistently distinguish corrupted inputs from originals.

Practical Implication

Builders evaluating vision systems should not read a scalar representation-similarity score as proof that image content is being integrated. Pair internal probes with controlled visual corruption and downstream task evidence.

Agent-Ready Context
Across **13 multimodal models** from five families, replacing visual tokens with Gaussian noise sharply reduced accuracy. Yet **four standard alignment measures** did not consistently distinguish corrupted inputs from originals.

Builders evaluating vision systems should not read a scalar representation-similarity score as proof that image content is being integrated. Pair internal probes with controlled visual corruption and downstream task evidence.

The proposed **principal-angle gap** tracked accuracy more consistently under graded corruption, but structured irrelevant images still exposed cases where geometry and performance diverged. It remains a diagnostic, not a direct content-understanding score.
Connected Context · Feed7 Judgment

This narrows what mechanistic alignment scores can establish for vision systems: similar internal representations may persist even when useful visual information is destroyed. It reinforces the prior case for controlled stress tests and behavior-level diagnostics, while showing that the proposed geometric replacement also needs task-grounded validation because irrelevant structured images can separate geometry from performance.

SABRE: Scalable and Automated Benchmarking of VLMs under StressSABRE provides the kind of controlled visual stress testing this Signal says must accompany internal similarity measures, so representation claims can be checked against downstream behavior.DiaVLo: Diagnosing Behaviours of Vision-Language ModelsDiaVLo’s causal concept diagnostics complement the principal-angle probe, but both remain dependent on behavioral specifications and cannot substitute for evidence that the visual content affects task performance.The Low Frequency Trap: Video Language Models Fail at Simple Event BookkeepingBoth expose apparently improved aggregate signals that do not prove faithful use of the underlying input: final counts can hide missed events, just as alignment scores can hide corrupted vision.A Lie Detector Test for Language Models: Reading Knowledge a Model Won't RevealPIR supports looking inside models when outputs are ambiguous, while this Signal supplies the corresponding limitation: an internal probe itself must be validated against controlled interventions and task outcomes.
Context Map
benchmarkimage#agent-evals#benchmark-integrity
Uncertainty
The proposed **principal-angle gap** tracked accuracy more consistently under graded corruption, but structured irrelevant images still exposed cases where geometry and performance diverged. It remains a diagnostic, not a direct content-understanding score.