THE GRAPH REMEMBERS: EXPLORATION COLLAPSE AS THE DOMINANT FAILURE MODE IN INTERACTIVE REASONING AGENTS, AND A TRAINING-FREE CORRECTIVE

Interactive-reasoning agents fail ARC-AGI-3 in a way frontier capability does not fix: they collapse into repetition. We measured it. Across our logged runs, a bare LLM click-agent (Vanguard-style) produced a repeat-ratio of 0.02 - effectively two candidate actions per hundred - including one trace of 46 consecutive identical clicks on a single coordinate. A state-graph-driven explorer over identical games measured 0.68+. We present the diagnosis, a training-free corrective (explicit state graph as action authority; the LLM demoted to advisor, consulted only at frontier exhaustion), and first-board results: submission 55661318 scores 0.05 on the private eval with four embedded hash-verified public plans - low in absolute terms, on a board whose leader holds 3.57 of a theoretical ~400.