TLDR: A fixed dataset quietly decides what a model gets to practice, so the next step in recursive self-improvement is letting models generate their own curriculum. Absolute Zero grounds proposed coding tasks in code execution, R-Zero pairs a Challenger with a Solver that grades itself by majority vote, and R-Few adds a small supply of trusted human examples to keep the loop from drifting. The catch is that difficulty is an imperfect proxy for progress: a verifier only establishes what its tests cover, and agreement among sampled answers can repeat the same mistake. We cover why the dataset has been making decisions for us, what execution can and cannot verify through deduction, abduction, and induction, why pseudo-label quality declines as questions get harder, and what POET teaches about stepping stones. We close with the evidence a self-generated curriculum would need to show: progress across multiple rounds, under a fixed budget, on evaluation tasks the model cannot influence. A...
learn more
