This paper proposes EvoCurr, a self-evolving curriculum learning framework for complex problem solving. EvoCurr adapts the learning progress of the solver by generating a sequence of problem instances with increasing difficulty. When the solver encounters difficulty, the difficulty is lowered, and when the solver successfully solves, the difficulty is increased, maintaining an optimal learning path. The solver, implemented as a code-generation model that generates Python decision tree scripts, gradually acquires the skills necessary for complex decision-making tasks. Experimental results demonstrate that the proposed method significantly improves task success rates and solution efficiency compared to existing direct solution methods.
