Anyone figured out how to restore Markdown formatting on the o3/o4-mini models released today?
For o1, the docs pointed to prepending “Formatting re-enabled” to the Developer message.
Despite prepending “Formatting re-enabled\n” at the beginning of the request, o3 and o4-mini output flat text.
Example below:
This seems a regression in behavior? Or maybe there’s a different way to turn formatting back on?
These AI models have been busting the Playground for a while also.
I first reversed your screenshots back to an inferred question.
Then I go after the one that looks more of a challenge: o3.
I give it the most logical instructional lines for a system message. How’d it do?
It’s still the AI’s own creative use, but works okay.
I’m running markdown2 and MathJax, and a balancing act to keep then perfect…
Something at the end of the system message, where the AI has post-training on schemas:
# Responses
## CommonMark
- Your output is CommonMark 0.31.2 complant markdown format, required by your response rendering environment.
- Code output: fenced and typed markdown code block
- LaTeX formula output: Enabled for inline and block
- headings and bullets and list nesting follow the very style shown right here
You can probably mandate hyphen-style bullets distinctly on the AI if you want. The model also all-but-watermarks its output with long dashes and other Unicode.
Hmm. Thanks for the help, but let’s see if there’s an official response, since I’m quoting the official OpenAI doc for re-enabling markdown for o1.
The [official doc](https://platform.openai.com/docs/guides/reasoning-best-practices#how-to-
prompt-reasoning-models-effectively) says:
o1-2024-12-17, reasoning models in the API will avoid generating responses with markdown formatting. To signal to the model when you do want markdown formatting in the response, include the string Formatting re-enabled on the first line of your developer message.Any official guidance will be appreciated.
Any “official answer” is no answer. I tried.
We can just have gpt-4.5-preview be the judge of what o4-mini is producing when switching models on a context. Not compliant.
The official word is ChatGPT will help you (in Korean also for some reason) because it does make headers using hashes…but won’t actually help you except in terms of a writing composition.
