Attention Convergence Across Architectures: Cross-Lingual Structure and Surrogate-Dependent Causal Effects

Transformer-based LLMs exhibit diverse attention organizations, yet cross-architecture comparisons remain scarce, and the functional role of observed patterns is rarely tested causally. We combine observational analysis with four-mode attention replacement across 11 model configurations from five families (Qwen3 0.6B–14B, Llama-3.1-8B and its base variant, Mistral-7B, InternLM3-8B, and the 47B MoE Mixtral-8x7B) in Chinese and English, all rerun under a single protocol. Observationally, over 90% of layer–head units fall in the converged or focused regimes, yet terminal-layer consistency plateaus at 0.70–0.79 (on canonical inputs) rather than unity. Convergence depth is architecture-specific, with skeleton stabilization and positional-attractor formation following different orderings; and a conserved sharp→diffuse→sharp entropy schedule holds across all observationally tested architectures. Causally, under prefix-smoothing surrogate interventions (damage depth ordering is surrogate-dependent; Appendix T), early-layer (first 25%) interventions cause the largest generation-perplexity degradation in 42 of 44 deterministic mode–language comparisons (11 configurations × 2 deterministic modes × 2 languages), including the MoE model, even where binary probe accuracy saturates; probe-level causal-order effects exceed one step in at most one of 33 configuration–band cells, and the standard uniform-causal/uniform-global contrast carries a support-size confound that we make explicit; and cross-lingual fragility is depth-structured: Chinese is directionally more fragile under early-band intervention in 9 of 11 configurations on the free-form perplexity track (metric-dependent; Appendix R), mid layers split the architecture families, and the late band returns to symmetry in 5 of 11. To our knowledge, these results yield the first systematic cross-architecture atlas of attention convergence under a unified protocol, with testable implications for KV-cache eviction scheduling—and a caution for interpretability practice: which layer appears causally critical depends on how attention is intervened upon.