MIRAGE reframes VLM hallucinations as reconstruction

The strongest takeaway is not that VLMs are “seeing” without vision, but that their learned priors are strong enough to support surprisingly coherent reconstruction. That’s useful as a safety warning and as a clue that benchmark performance may be inflated by text-based shortcuts.

• –The paper highlights a real evaluation problem: models can score above chance, or even strongly, without actual image input, so benchmark numbers may overstate visual grounding.
• –The mirage-vs-guessing gap suggests prompt framing changes inference depth, but the “geometric reconstruction” interpretation is still an inference, not a proven mechanism.
• –The most practical implication is for medical and high-stakes multimodal systems: developers need counterfactual tests that verify the model actually used the image.
• –The commentary is directionally right that internal representations matter, but it likely overreaches when it treats hallucination as evidence of “deeper” understanding rather than a mix of priors, dataset leakage, and textual pattern completion.
• –As a research result, MIRAGE is more compelling as a benchmark/safety critique than as proof of a new theory of cognition.