DeepConf: Scaling LLM reasoning with confidence, not just compute

GEPA: Reflective prompt evolution can outperform reinforcement learning