Simin Yuan

Mechanistic and reliability research on large language models. I study self-referential loops and constraint recovery — a controlled study showing that the effect of scale reverses: at 1.5B a volume account holds, at 7B it fails entirely, and only coupling restores recovery (26.7%, p=.0078). Accepted at NeurIPS 2026 workshops (NewInML @ Paris; EvoRobust @ Sydney). My second thread is whether verification itself can be trusted — precheck, greencheck and self-auditing-agent ask whether a green build passed because the gates caught something, or because they never ran. Papers, raw trials and reproduction code: github.com/simin-yuan