Can an AI get smarter by playing against a rival that keeps raising the difficulty, instead of just fixing its own mistakes?

This explores how training a model against a second role (a challenger, critic, or problem-setter) differs from training one model to critique and fix its own work. The main difference is where the pressure comes from. In self-play, one side's job is to make the other's life harder, so the game itself manufactures difficulty and feedback. In Can language models improve themselves without any external training data?, a proposer writes problems calibrated to the solver, and the solver learns from majority-vote agreement, with no human labels or ground-truth answers. In Can language models learn skills without human supervision?, a Challenger raises difficulty as a curriculum while a neutral Judge gives binary verdicts that serve as the reward.

Adversarial setups can also stand in for the verifier. Can adversarial critics replace task-specific verifiers for reasoning? trains a critic to tell expert answers from the policy's, so the policy improves by getting harder to distinguish. That gives it reasoning RL on tasks like poetry writing, where no checker exists. It does lean on expert demonstrations, so it is adversarial rather than fully self-supplied. Single-model self-revision has no opponent. Can models learn to evaluate their own work during training? trains a model to assess its own output in the unused space after the answer, so it internalizes reward computation at zero inference cost. That is cheap, but the model is grading itself.

Self-revision's main training problem is practicing on the right mistakes. Why does self-correction training on offline data fail? finds that fine-tuning on pre-collected correction examples fails, because the errors in the data aren't the errors the model makes at test time, and the model collapses into one correction habit. What works is multi-turn online RL, where the model corrects its own live mistakes. Self-play's main problem is different: keeping the game from spiraling. The Challenger-Judge loop only works with a generalization safeguard balancing the adversarial pressure. Do overly hard RLVR samples actually harm model capabilities? shows that near-impossible problems teach shortcuts instead of reasoning, which is a warning for any challenger that keeps escalating. Can human data steer self-play RL toward human-compatible behavior? points to a related risk: a game optimizes for beating the opponent, not for what people expect. In that work, a small amount of human data steered self-play toward human-compatible behavior.

The two ideas also combine. Ctx2Skill's self-play evolves skills through natural-language edits, which is self-revision inside an adversarial loop. And Does constraining edits make skill learning more stable? shows that unconstrained self-rewriting is unstable. Its fix is a held-out validation gate that can veto edits, plus a buffer of rejected edits kept as negative feedback. That is essentially a second party checking the first. So self-play is best seen as a source of outside pressure, and self-revision as a way of changing the model or its instructions. Revision without pressure has to smuggle in some check of its own.

The corpus has no head-to-head comparison of the two on the same tasks. What it does show is that the adversarial versions mostly solve the missing-feedback problem, while the self-revision versions mostly solve the wrong-kind-of-practice problem.