[derefr]

When you say "improve an svg like this", how are you imagining setting that workflow up? Are you just feeding them the SVG to iterate on; or are you giving them access to a browser to look at the rendering of the SVG?

I ask because:

Insofar as the original pelican test is zero-shot, it effectively serves as a way to test for the presence of a kind of "visual imagination" component within the layers of the model, that the model would internally "paint" an SVG [or PostScript, etc] encoding of an image onto, to then extract effective features from, analyze for fitness as a solution to a stated request, etc.

But if you're trying to do a multi-shot pelican, then just feeding back in the SVG produced in the previous attempt, really doesn't correspond to any interesting human capability. Humans can't take an SVG of a pelican and iteratively improve upon it just based on our imagined version of how that SVG renders, either! Rather, a human, given the pelican, would simply load the pelican SVG in a browser; look at the browser's rendering of the pelican; note the things wrong with that rendering; and then edit the SVG to hopefully fix those flaws (and repeat.)

I imagine current (mult-modal and/or computer-use) LLMs would actually be very good at such an "iterative rendered pelican" test.

[irthomasthomas]

I'm talking about two type of improvement, model improving, and prompt based improving. I am noticing that the baseline output has a lot more going on, the model has improved, yet it still makes those obvious looking mistakes with the shape of the frame or disconnected limbs etc.

And I am saying that if you take one of these SVGs and ask an LLM to look for flaws, it rarely spots those obvious flaws and instead suggests adding a sunset and fish in the birds mouth.

[tskj]

This is also my gripe with a lot of this stuff, always evaluating models on what they can literally oneshot is completely pointless; it's not how anything works, neither for humans nor for scaffolded AIs. I guess it's neat if you want to argue that a certain level of intelligence can "never be achieved" in a single forward pass, but like, so what. No one cares about that, except people who have already decided to be anti AI.

(not that I am in any sense pro AI, but it's just a weird lack of intellectual rigor)

[irthomasthomas]

Asking a model to improve its output is not one-shotting tho? My observation was that asking an llm to iterate and improve a response causes it to add more stuff, rather tha repair the broken stuff. And that model progress in general has the same pattern. This new model adds more details to its responses but continues to make mistakes at about the same rate.

[losvedir]

The question was whether you were giving it the rendered image and using the model's visual modal capability, or feeding back in the textual SVG.

It's hard to "imagine" what the rendered SVG looks like, for both humans and LLMs, so just iterating on text won't really be as useful of a test. But if you show it what it rendered, it might observe the bad-looking bicycle and be able to fix the text that way.

[irthomasthomas]

"I've even experimented with feeding the broken pelican svgs to an image model to look for flaws, and they still fail to spot the broken elements."