Hacker News new | ask | show | jobs
by BoorishBears 210 days ago
This shows some gaps in the "same prompt to every model" approach to benchmarking models.

I get that it's allows ensuring you're testing the model capabilities vs prompts, but most models are being post-trained with very different formats of prompting.

I use Seedream in production so I was a little suspicious of the gap: I passed Bytedance's official prompting guide, OPs prompt, and your feedback to Claude Opus 4.5 and got this prompt to create a new image:

> A partially eaten chicken burrito with a bite taken out, revealing the fillings inside: shredded cheese, sour cream, guacamole, shredded lettuce, salsa, and pinto beans all visible in the cross-section of the burrito. Flour tortilla with grill marks. Taken with a cheap Android phone camera under harsh cafeteria lighting. Compostable paper plate, plastic fork, messy table. Casual unedited snapshot, slightly overexposed, flat colors.

Then I generated with n=4 and the 'standard' prompt expansion setting for Seedream 4.0 Text To Image:

https://imgur.com/a/lxKyvlm

They're still not perfect (it's not adhering to the fillings being inside for example) but it's massively better than OP's result

Shows that a) random chance plays a big part, so you want more than 1 sample and b) you don't have to "cheat" by spending massive amounts of time hand-iterating on a single prompt either to get a better result

2 comments

100%. Between tuning prompt variations depending on the model and allowing a minimum number of re-rolls, this is why it takes a while to publish results from the newest models on my GenAI comparison site.

Including a "total rolls" is a very valuable metric since it helps indicate how steerable the model is.

not adhering to the prompt guide is def a valid strong criticism. resampling i think less so for the demo just because fewer people look at k samples per model, so just taking literally the first one has the fewest of my own biases injected into it
I actually think it's ok to inject your own bias here: if you're deploying these models in production, then you probably test on your own domain other than half eaten burritos lol

But individual users usually iterate/pick, so just sharing a blurb about your preference is probably enough if you choose 1 of n