Theseus Labs and Partners Release RSI Roadmap Report with HCI Metric and L1-L5 Autonomy Levels

Theseus Labs, together with Tsinghua University, ByteDance, ModelBest and other research and industry partners, released a report titled "The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement" in September 2026. The report introduces the "improvement loop" as an analytical unit, systematically reviewing technical routes, industrial practices and evaluation challenges of recursive self-improvement (RSI). It also proposes the Headroom-Closed Index (HCI) and a five-level autonomy roadmap from L1 to L5.

According to the report, frontier mathematics has reached an HCI of 86.4 by 2026, while areas such as legal reasoning and multimodal reasoning still lag. Major AI companies including OpenAI and Anthropic are mostly at L1-L2, with some systems moving toward L3. The report highlights that fields with larger capability gaps, such as software engineering and tool-using agents, may benefit most from RSI.

The report also emphasizes that genuine recursive improvement should be measurable, maintainable, transferable and reversible, and that humans should continue to oversee the overall goals and safety boundaries.