TopK Sparse Autoencoders Across Three Model Architectures: Dictionary Collapse, Dense-Feature Degeneracy, and the Limits of Activation-Pattern Feature Matching

Version 2 (correction). Version 1 stated that no study of this kind can report more than about 8% same-seed SAE feature agreement, and restated it as a general bound. That universal claim is withdrawn. Paulo & Belrose (arXiv:2501.16615) report 30% for 131K-latent SAEs on Llama-3-8B under Hungarian matching at cosine > 0.7, and 42% under a maximum-cosine criterion; Braun et al. (arXiv:2405.12241) report above 90% for GPT-2 with ReLU SAEs. Neither was cited in v1. The 8% measurement is unchanged: it is the ceiling of THIS procedure — activation-fingerprint similarity, reciprocal matching, a permutation-calibrated threshold, and a matchable-feature denominator — not a bound on seed stability in general. The cross-architecture result is unaffected, being a within-instrument comparison of 0.18–0.33% against that same 8% ceiling.TopK SAEs trained at matched sparsity and dictionary size on Llama-3.2-3B, Mistral-7B-v0.3 and Qwen2.5-3B. Reports abrupt dictionary collapse with partial recovery, dense-feature degeneracy invalidating frequency-based feature selection, and a null cross-architecture matching result whose detection floor is measured by planted-signal power analysis (MDES 0.95-1.00).