Volume or Coupling? A Scale-Dependent Dissociation in Constraint Recovery of Language-Model Loops

Recent work on self-referential large-language-model (LLM) loops suggests that isolated recursive loops degrade toward closed attractor states and that exchange with an independent process supports recovery. We study the mechanism behind this effect using a minimal perturbation–recovery benchmark: an instruction-tuned model maintains an explicit output constraint, a conflicting instruction is injected, and recovery is evaluated across coupled dual-agent, single-agent, volume-matched single-agent, and four-fold-volume conditions. At 1.5B parameters, the apparent advantage of coupling is largely explained by generation volume: coupled recovery is 43%, while a volume-matched single-agent condition reaches 37% (C vs. X, p = .774). At 3B, all conditions approach ceiling performance, limiting discrimination. At 7B, however, single-agent and volume-matched conditions recover in 0/30 runs, while coupling recovers in 8/30 runs (C vs. X, p = .0078). Exact trajectory replays show that the directly perturbed agent remains deeply captured, while a partner agent outside the perturbation history provides constraint-consistent evidence and enables recovery. These results suggest that robust agent recovery depends not only on additional generation resources but also on structural separation from failure-inducing context. The study provides a controlled benchmark for evaluating robustness mechanisms in interactive and multi-agent AI systems.