AI startup autoformalizes mathematical proof to accelerate alignment research
Resolution uses AI agents to formalize 600k lines of Lean code for singular learning theory, spending an estimated $135,000 in compute to reduce weeks-long manual work to days.
1 source · cross-referenced
- AI safety startup Resolution autoformalized the resolution of singularities in statistical learning theory using AI agents, producing 600k lines of Lean code in about a week.
- The work cost an estimated $135,000 in token spend and relied on an 112-page atomic plan executed by orchestrated agents.
- Resolution says this ‘lays the groundwork’ for automating research into loss landscape geometry and generalization.
- The company frames this as part of an ‘industrialization of mathematics’ that could enable a new, rigorous science of research automation.
AI safety startup Resolution (formerly Sequent) reports it has automatically formalized mathematical statements related to statistical learning theory using AI agents. The team autoformalized the resolution of singularities, a result they say ‘lays the groundwork for further automated research into the relationship between loss landscape geometry and generalization.’
The effort began with a researcher drafting a high-level plan and a detailed 112-page plan broken into atomic steps, then using an orchestrator agent to delegate tasks to specialized agents for pen-and-paper math, Lean code generation, and progress tracking. According to Resolution, the execution phase required minimal additional scaffolding beyond the initial plan and previously distilled agent skills.
Resolution estimates the work took about a week and consumed on the order of 100 billion input tokens (90% cache reads) and about 300 million output tokens, totaling an estimated $135,000 in token spend. The result was about 600,000 lines of Lean code.
Resolution states it has already seen benefits from the autoformalized library, including extending parts of singular learning theory and enabling automation of theory components to better understand sampling, which informs how models learn. The company frames this as part of an ‘industrialization of mathematics’ that could enable a new, rigorous science of research automation.
- Nov 11, 2026 · Wired
EU proposes Digital Fairness Act to ban dark patterns and restrict addictive design features
Trust75 - Nov 11, 2026 · AI Snake Oil — Narayanan & Kapoor
New nonprofit STABL launches to study societal bottlenecks of advanced AI
Trust79 - Nov 10, 2026 · TechCrunch — AI
New York enforces AI safety law requiring registration and incident reporting for frontier labs
Trust75