Killing Coding Agent Slop With Adversarial Self-Play

Generation is a fundamentally hard and unforgiving search problem. Verification is simpler. OpenAI studied this asymmetry in their work on prover-verifier games, where they explicitly used weaker models as verifiers for math proofs.

Their italics. This framing is funny to me because the two options presented are:

• Stronger AI generates code, human verifies it: Suboptimal because AI can generate slop faster than humans can review it.
• Stronger AI generates code, weaker AI verifies it: Faster and better than no review at all.

There's a missing third option: a human generates the code and an AI reviews it. Humans are arguably still stronger than the strongest AI models, so according to their thesis, wouldn't you get an even higher quality result by having humans write the code and AI clean it up?

I presume the answer is yes, but that if you're balancing quality against the costs of time and effort, you choose the good-enough options, and AI/AI was good enough for their benchmark tasks.

Happy to see this, it seems to me that as long as we have this assymetry we should be able to performs this bootstrap method.

I don't think our current models, and their setup, can generate synthetic software that is slop-free. And there are, at least to me, unknown metrics for how much and sloppy slop a codebase can handle before diverging irrecoverably.