Mirror Horizon Measures AI Reflection and Viable Continuations.

This paper introduces "Mirror Horizon," a concept derived from Mirror Theory, which proposes evaluating intelligent systems not just by their representations but by their ability to sustain coherent continuations through repeated reflection. The operational measure for this is "viable path entropy" (VPE), which quantifies verified continuation capacity under a finite computational budget. VPE breaks down bounded capability into two components: the probability of achieving a viable continuation and the diversity of verified continuation modes reached. The theoretical framework posits intuition as local constraints, taste as invariant-selecting pressure, and reflection as taste-guided resolution of underdetermination, with geometry as the learned structure for stable future reflection. In language model reasoning experiments using GSM8K, increasing the token budget significantly expanded verified reachability and diversity. Notably, a smaller model (Qwen2.5-1.5B) demonstrated a stronger mirror horizon than a larger one (Qwen2.5-3B) under specific reflection protocols, indicating that capability is about accessible verified continuation capacity rather than just parameter count or one-shot accuracy.