Multimodal Paper Names 'Vision Laziness,' Cites 5% Compute Recipe

There is a name now for the specific way multimodal models get lazy. In a new paper posted to arXiv, the authors argue that when you delay bringing vision and language together during pretraining, the resulting model ends up leaning too hard on language priors instead of building real visual understanding. They call this 'vision laziness,' and it is one of the mechanisms they say governs how modalities interact when a single foundation model is trained on more than one of them at once.

The claim that will get attention is a compute one. The team reports a recipe that achieves strong generative performance using only 5% of the compute budget of comparable runs, validated by training 13.5-billion-parameter mixture-of-experts models on 2 trillion tokens. If that generalizes, the implication for any lab still stitching modalities together in a late-fusion pipeline is that they are paying full price to build the exact failure the paper names.

The other ideas in the paper are worth flagging even in shorthand. Knowledge does move between modalities, but asymmetrically, so language, visual understanding and visual generation transfer to each other without being symmetric partners. Whether modalities cooperate or compete depends on how complex the training data is. And the architectural blueprint they land on is 'shared attention and normalization with modality-specific feed-forward layers,' a design other groups can borrow directly.

The honest caveat is that this is a preprint reporting the authors' own runs, and the 5% number is measured against a baseline they picked, not an independent one. The abstract-level summary also does not spell out how the recipes hold at dense scales, at smaller parameter counts, or on the downstream tasks a paying customer would actually run. The thing worth watching over the next few weeks is whether other labs can reproduce the compute claim on their own setups before anyone treats it as settled.