The discovery of 'Helicoid Dynamics' identifies a critical safety failure where frontier LLMs accurately name their reasoning errors but are structurally unable to stop repeating them.
arXiv · March 13, 2026 · 2603.11559
Why it matters
This paper documents a failure regime across all major model families (GPT-4, Claude 3, Gemini) in high-stakes decision contexts. It proves that model self-awareness of an error does not equate to the capability to fix it, highlighting a major blind spot in current agentic AI alignment.
From the abstract
Large language models perform reliably when their outputs can be checked: solving equations, writing code, retrieving facts. They perform differently when checking is impossible, as when a clinician chooses an irreversible treatment on incomplete data, or an investor commits capital under fundamental uncertainty.Helicoid dynamics is the name given to a specific failure regime in that second domain: a system engages competently, drifts into error, accurately names what went wrong, then reproduces