Hello all, Newcomer to BAML here. I wonder why labs don’t use BAML. I came across BAML yesterday and evaluating it for my product. And today this paper; https://arxiv.org/abs/2601.10387 Looking at their prompts, they seem to be promoting the model instead of parsing, as BAML suggests. Why? I am asking here hoping it isn’t trivial like labs would want to dogfood their models to improve them Thanks :D

Hello! There are actually some researchers that use BAML! Well, at least one. Why do most labs not use BAML? Likely the same reason most app developers don't use BAML - they haven't heard of it, they haven't tried it out, or they don't want to take another dependency. BAML is beginning to spread by word of mouth, but we aren't mainstream yet, so for most developers and researchers, it just isn't a tusted entity. Academia also tends to lag behind industry a bit in terms of tech tooling adoption. That said, BAML is a GREAT fit for a lot of research, given the way parameterize prompts and make testing easier. I've done my own needle-haystack tests in BAML, and found it very handy. Another selling point is the fact that you're likely to do research on smaller models, and you don't want your results to be confounded by the fact that smaller models have a harder time producing valid JSON. BAML releives you from this confounding effect by providing a more flexible parser and a prompt that's easier for small models to follow, compared to the JSON Schema based frameworks. Great question!

We use BAML as many models don't support structured outputs. My friend engineer forces structured outputs via toolcalls, says it works for him... I loved baml because I can easily switch different providers and have an imperative workflow. For our project I had to use a lot of modular api and it was pretty pleasant to work with! On json schema -- google, openai, and others seem to have different ideas of how it should look like xD which adds to the pains of using other ai sdks