A Named Answer Is Not a Correction: The Moral Self-Correction Threshold Moved From 22B to 3.8B on Benchmarks a Context-Blind Policy Can Pass, and the Controls That Would Separate Steering From Correcting

APA

Mahendrakar, P. (2026). A Named Answer Is Not a Correction: The Moral Self-Correction Threshold Moved From 22B to 3.8B on Benchmarks a Context-Blind Policy Can Pass, and the Controls That Would Separate Steering From Correcting [Preprint]. Life of Research. https://doi.org/10.5281/zenodo.23443486

Copy APA

BibTeX

@misc{lor1142026,
  title        = {A Named Answer Is Not a Correction: The Moral Self-Correction Threshold Moved From 22B to 3.8B on Benchmarks a Context-Blind Policy Can Pass, and the Controls That Would Separate Steering From Correcting},
  author       = {Pranay Mahendrakar},
  year         = {2026},
  publisher    = {Life of Research},
  doi          = {10.5281/zenodo.23443486},
  url          = {https://lifeofresearch.com/preprints/114},
  keywords     = {moral self-correction, social bias evaluation, BBQ, instruction following, emergent abilities, abstention, fairness benchmarks, evaluation methodology},
  note         = {LOR-00114}
}

Copy BibTeX More on the record page