ARIADNE: Persistent Structural Anchors for Multi-Turn Coherence in Large Language Models

Large language models lose the structure of an idea after one or two divergent exchanges, and they cannot reliably return to it. This proposal names the failure, explains its cause, and specifies an architecture to fix it: ARIADNE. ARIADNE adds three things to the generation loop: persistent structural anchors (compact, machine-readable skeletons of each turn, stored as a tree and visible only to the model); an explode–recoil cycle (the model branches into candidate answers, then prunes drift against the anchors before anything is shown); and addressable return (context for the next turn is assembled from the anchor chain, not from the raw transcript). The proposal includes an evaluation benchmark, Divergence-and-Return, that any lab can run to measure the problem and the fix. It is offered openly to research labs.