Written Is Not Kept: What a Single-Layer Readout of the Residual Stream Licenses When Adjacent Layers Undo Part of Each Other's Writes, Why Cancellation and Iterative Refinement Are Not Rival Findings, and the Persistence Check Interpretability Claims Skip

APA

Mahendrakar, P. (2026). Written Is Not Kept: What a Single-Layer Readout of the Residual Stream Licenses When Adjacent Layers Undo Part of Each Other's Writes, Why Cancellation and Iterative Refinement Are Not Rival Findings, and the Persistence Check Interpretability Claims Skip [Preprint]. Life of Research. https://doi.org/10.5281/zenodo.23838677

Copy APA

BibTeX

@misc{lor1732026,
  title        = {Written Is Not Kept: What a Single-Layer Readout of the Residual Stream Licenses When Adjacent Layers Undo Part of Each Other's Writes, Why Cancellation and Iterative Refinement Are Not Rival Findings, and the Persistence Check Interpretability Claims Skip},
  author       = {Pranay Mahendrakar},
  year         = {2026},
  publisher    = {Life of Research},
  doi          = {10.5281/zenodo.23838677},
  url          = {https://lifeofresearch.com/preprints/173},
  keywords     = {mechanistic interpretability, residual stream, sparse autoencoders, layer correction, logit lens, activation steering, transcoders, iterative inference},
  note         = {LOR-00173}
}

Copy BibTeX More on the record page