Post

Grace
@gracekind.net

There’s some cool recent research on this phenomenon! It turns out vision language models excel at image benchmarks *even when the actual images aren’t provided,* because the answers are implicit in the questions! arxiv.org/abs/2603.21687

arxiv.orgMIRAGE: The Illusion of Visual UnderstandingMultimodal AI systems have achieved remarkable performance across a broad range of real-world tasks, yet the mechanisms underlying visual-language reasoning remain surprisingly poorly understood. We r...
Kyle L@haikugoat.bsky.social

It's really instructive to look at an LLM's output and piece by piece look to see how much of it was already implicit in the prompt, and where exactly. It makes me realize how much I'm telling on myself all the time.

51 reposts233 likes

👻 spoOoOoky frood 💀
@huwupy.kawaii.social

Now flashing back to one of those old ML models that would never, ever tell you there were zero giraffes in a picture, because nowhere in its training data had it been asked that about a picture with no giraffes

norvid_studies
@norvid-studies.bsky.social

a similar thing happened to me except I got the question wrong bsky.app/profile/norv...

norvid_studies@norvid-studies.bsky.social

oh I thought this notification was re: the "turning all the simcluster art videos into one syncretic video" thread comment. premove blunder, ?? evaluation losing by force... I'll keep it up though, no takebacks in horseshoes or hand bisking... @abeliansoup.bsky.social

affine
@refinement.systems

"(...) by generating a reasoning trace, indistinguishable from a correct one, solely based on a described mirage rather than a real input, while at the same time scoring high on the benchmarks. Indeed, these models outperform not only other AI models, but in some cases also human experts." wtf

cl0q
@cl0qsearch.bsky.social

The benchmark lesson here is brutal — if models can ace visual QA without looking at the image, the benchmark measures language priors, not vision. Every multimodal eval needs a 'blank image' baseline.