Efficient post-training is targeting both supervision and discovery

Uncertainty-Guided Progressive Contraction claims full-SFT-level reasoning performance with only 0.22% of the optimization budget, while diverse initialization improves LLM-driven discovery across five tasks and 12 harnesses. CURIO adds curiosity-driven test-time learning that improves discovery and denoising while increasing program diversity; SALD extends teacher-free self-improvement ideas to diffusion models, but lacks quantitative evidence. These methods remain promising but need broader benchmarks, compute accounting, contamination controls, and independent replication.