Already 8 months old when the scout caught it.
Vibe-proving
What it is
Vibe-proving is the mathematical-research analogue of vibe coding: a mathematician drives a large language model through an iterative generate-referee-repair loop — propose a proof step, have it checked, patch what's wrong, repeat — to work toward genuinely unproven results, rather than trusting a single model output outright.
Why it's emerging now
Reasoning models with much larger inference-time budgets (Gemini Deep Think, and GPT-class "thinking" modes) started producing real partial progress on open problems in late 2025, not just retrieving known results. Practitioners needed a name for the resulting workflow, and it has since been picked up by independent research (an arXiv case study resolving an open spectral-region conjecture with a documented generate-referee-repair pipeline) and by major labs building agents explicitly around the loop, like Google DeepMind's Aletheia, whose "Advisor" collaboration model is built on the same iterative cycle.
How it relates to what you know
It's a domain-specific instance of the generate-verify-repair loop you already know from agentic coding and tool use, applied to formal or informal mathematical reasoning instead of software. It sits next to reasoning models / test-time compute on your ledger — vibe-proving is what that extra inference-time compute gets spent on in a research-math setting, with a human or formal verifier playing the same "referee" role that evals and guardrails play for agentic coding.