When the Picture Stopped Getting Through, Four Standard Scores Barely Moved

Ask a multimodal AI a question about a photograph and it answers. Watch its insides while it works, and you can see the model's representation of the image and its representation of the words converge, layer after layer, until in the deepest layers they very nearly coincide. For several years that convergence has been read as the model folding what it sees into what it reads. It is the closest thing the field has to a picture of a machine looking.

So Hong-Han Wang, Yuntao Wang and Hu Ding, at the University of Science and Technology of China in Hefei, broke the picture and watched the meter. In a paper posted on 24 September 2026 and listed by its authors as accepted to the NeurIPS 2026 conference, not yet in the proceedings, they took 13 multimodal models from five families, ranging from 0.5 billion to 72 billion parameters, and replaced the visual information arriving at each model's language half with random noise. Answer accuracy collapsed. The convergence did not.

The field's word for that convergence is alignment, and on an AI page it is the wrong word in the reader's head: this has nothing to do with whether a system shares human values. Here alignment is pure geometry. It asks how much the space the picture tokens occupy inside the model overlaps with the space the word tokens occupy, and a high overlap has been taken to mean the two are genuinely interacting.

What the team corrupted matters, because it is not the image. A model of this kind runs in three stages: a vision encoder that looks, a projector that rewrites what it saw into tokens the language model can read, and the language model itself. Wang and colleagues left the camera end untouched and replaced the projector's output tokens with Gaussian noise, each one rescaled to the same magnitude as the token it replaced, so that content was removed and signal strength was not. On MMBench, a standard multiple-choice benchmark, accuracy dropped by 38 to 50 percentage points in every one of the 13 models. Whatever the models had been doing with the picture, they could no longer do it.

The scores did not follow. Four standard ways of measuring how far two sets of internal representations have converged went into the same run: CKA, SVCCA, MIR and the leading principal-angle cosine. Across the models, all four failed to consistently separate the corrupted stream from the original, and for three of them the median ran the wrong way, rating the noise-filled run as the better aligned of the two. The team calls this the alignment illusion.

The scores were reading the wiring

Where does a similarity that carries no content come from? Both streams travel through one stack of weights, and the authors trace the effect there. Measured before the language model, between the fixed projector output and the text, the similarity sits near what random vectors would give; the near-perfect values appear only after both have passed through the shared pathway. Bypassing the feed-forward block inside each layer disturbs it more than bypassing attention does, in every model tested. That block ends in a matrix that is not directionally neutral: its strongest output direction dominates the next by well more than a random matrix of the same shape would give. And the directions along which picture and text appear to align sit inside that matrix's dominant ones. A shared projection that pulls both streams toward the same axis will inflate the score with no coupling between the two inputs at all, a result they also prove as a bound. They call it weight-induced alignment.

None of this is unfamiliar in kind. In 2021 William Timkey and Marten van Schijndel showed that in text-only transformers a handful of rogue dimensions dominate standard similarity measures while mattering little to what the model actually does. What is new is the multimodal case, where two kinds of input are pushed through a single set of such weights.

A second reading of the same geometry

The authors' proposal is not a new instrument but a second reading of the one already in use. Principal angles between the two subspaces come as a spectrum, ordered from the closest direction downward, and the standard score keeps only the first number. They subtract the second from the first and call the difference the principal-angle gap. The reasoning is that the leading angle is the one the shared weights manufacture: if genuine visual structure is present the alignment spans several directions, the second angle stays high too, and the gap is small. Strip the content out and everything below the first collapses, so the gap opens. It ranked the clean stream above the corrupted one in all 13 models, where the four standard scores managed it in 3 to 5. Under a graded blend from real tokens to pure noise, it followed accuracy more closely than any of the scores it was set beside.

Where the new reading stops

Then comes the test that keeps this from being a story about a repaired metric. Instead of noise, each question got a real photograph belonging to a different question. The projector encodes it faithfully, and simple probes recover the wrong image's category from its tokens far above chance; the language model does not throw it away. Median accuracy went from 85.4% with the right image to 36.1% with the wrong one, slightly below the 39.0% the models managed on noise and about 5 percentage points below what they scored with the picture removed altogether. An irrelevant image is not ignored. It is processed, and processing it costs more than having nothing to look at.

The principal-angle gap, meanwhile, behaved exactly as designed and still could not answer the question that mattered. It placed the wrong image between the right one and the noise, because what it reads is structure in the visual stream, and structure was there. Accuracy did not follow that order. Internal geometry can show that structured visual content is reaching the model's answer; it cannot show that the content is any use for the question being asked. The finding is drawn from one benchmark, multiple-choice questions and designs that route a projector into a language model, with open-ended generation, video and document tasks named as untested. Within that, the authors' summary is plain: internal alignment and task performance "should therefore be reported together, but not read as the same kind of evidence."