structural-depth-hypothesis

Tag

Cards List
#structural-depth-hypothesis

Self-Training Doesn't Flatten Language -- It Restructures It: Surface Markers Amplify While Deep Syntax Dies

arXiv cs.CL · 2026-05-21 Cached

This paper presents evidence that self-training on language model outputs does not uniformly flatten language but restructures it, with surface markers (discourse connectives, hedges, em-dashes) increasing while deep syntactic structures (passives, subjunctives, parentheticals) collapse, formalized as the Structural Depth Hypothesis.

0 favorites 0 likes
← Back to home

Submit Feedback