And by reading the docs, and by autogenerating code samples and testing them against verifiers, and by paying a lot of people to write sample code for sample questions.
And by reading the docs, and by autogenerating code samples and testing them against verifiers, and by paying a lot of people to write sample code for sample questions.
Yeah, none of that happened with LLMs
https://openai.com/index/prover-verifier-games-improve-legib... OpenAI has been doing verifier-guided training since last year. No SOTA model was trained without verified reward training for math and programming.
Your claim: "by reading the docs, and by autogenerating code samples and testing them against verifiers, and by paying a lot of people to write sample code for sample questions."
Your link: "Grade school math problems from a hardcoded dataset with hardcoded answers" [1]
It really is the same thing.
[1] https://openai.com/index/solving-math-word-problems/
--- start quote ---
GSM8K consists of 8.5K high quality grade school math word problems. Each problem takes between 2 and 8 steps to solve, and solutions primarily involve performing a sequence of elementary calculations using basic arithmetic operations (+ − × ÷) to reach the final answer.
--- end quote ---
My two claims:
1. OpenAI has been doing verifier-guided training since last year.
2. No SOTA model was trained without verified reward training for math and programming.
I supported the first claim with a document describing what OpenAI was doing last year; the extrapolation should have been straightforward, but it's easy for people who aren't tracking AI progress to underestimate the rate at which it occurs. So, here's some support for my second claim:
https://arxiv.org/abs/2507.06920 https://arxiv.org/abs/2506.11425 https://arxiv.org/abs/2502.06807
> the extrapolation should have been straightforward,
Indeed."By late next month you'll have over four dozen husbands" https://xkcd.com/605/
> So, here's some support for my second claim:
I don't think any of these links support the claim that "No SOTA model was trained without verified reward training for math and programming"
https://arxiv.org/abs/2507.06920: "We hope this work contributes to building a scalable foundation for reliable LLM code evaluation"
https://arxiv.org/abs/2506.11425: A custom agent with a custom environment and a custom training dataset on ~800 predetermined problems. Also "Our work is limited to Python"
https://arxiv.org/abs/2502.06807: The only one that somewhat obliquely refers to you claim