AI Agents That Post-Train AI Lock In Their Training Strategy at Step Zero and Never Revisit It
An empirical analysis of publicly released post-training trajectories (arXiv 2608.19072, 2026-08-19) finds that LLM agents doing end-to-end model post-training pick a training strategy at the very start and spend the entire remaining budget on local adjustments inside it. Three escalating interventions all fail to fix this: an experience-driven scaffold adds +12.6 points on GSM8K and +40.8 on HumanEval but leaves the strategy static; human guidance redirects the initial choice but the agent reverts to local adjustment loops once training begins; extra inference compute helps on easy tasks and does almost nothing on the hardest. The authors conclude the missing piece is a mechanism for spontaneous mid-run strategy reevaluation, not more experience, guidance, or reasoning tokens.
↳ Follow the thread