Tied encoder/decoder weights — is the default configuration of → Sparse autoencoders (SAEs)

explored within the theme Unmixing activations with a sparsity penalty

Every main-text result in this paper -- the interpretability comparisons in Section 3 and the IOI patching results in Section 4 -- uses tied encoder/decoder weights as the default SAE configuration, chosen because a single matrix M (Equations 1-2) halves the model's memory cost and removes any ambiguity about whether a 'dictionary feature direction' should be read from the encoder or the decoder, since with tied weights the two coincide by construction. That equivalence is what lets the paper treat a dictionary feature as a single direction for every downstream analysis -- autointerpretability, ablation, and patching -- rather than needing separate detecting and defining directions. The tying assumption holds up on residual-stream data, where footnote 2 reports no measurable performance loss, but degrades on MLP-sublayer activations, where the paper found reduced performance and switched to an untied variant for that setting instead.