Large to Small Model Stitching Destroys What You Built It to Carry Along the Fit.

An independent researcher posted early-stage work on model stitching, reporting that stitching large models to small ones destroys the properties the stitching was meant to preserve. The post requests feedback on whether the effect holds at realistic scale and on related prior work.

Why it matters: Model stitching is used in AI safety work to reuse sparse autoencoders and probes across models, so an effect that breaks the stitched connection could undermine that reuse.

Written by AI from LessWrong's published text.