Researchers measure concept transmission limits across scaling language models

The paper's core contribution isn't just a new metric, but evidence that semantic transmission between models correlates with human annotation error. This suggests a measurable bridge between what models learn internally and what humans actually care about evaluating.

This work sits alongside recent mechanistic studies of how models encode and transfer information. The prompt optimization algorithm here (14% improvement) echoes the trade-off patterns documented in 'Do Prompt Optimizers Break Requirements' from October 9th, where automated refinement improved one metric while degrading unscored ones. More broadly, this connects to the October 13th work on prompt adaptation mechanisms, which showed that different optimization techniques reshape internal computation differently. If articulation bounds reveal fundamental limits on what language can encode, that constrains how much semantic richness any prompt technique can actually extract, regardless of optimization method.

If the transmission metric holds across held-out model pairs trained on different data distributions (not just same-family scaling), that validates the framework as a genuine semantic measure rather than an artifact of training similarity. If not, watch whether the correlation to human error persists only within model families, which would suggest the metric captures training-time alignment rather than abstract concept transfer.