Transformer MLP Gate Thresholds Are Couplings to a Carried Reference Direction
arXiv:2609.31956v1 Announce Type: new Abstract: The corpus-mean direction of a transformer's residual stream is a component shared across all inputs, and is commonly removed by mean-centering before representational analysis. We present evidence that it is a functional component: the reference agai…
Read the full story at arXiv cs.CL ↗
Timeline · 1 report
- 2026-09-29 04:00 · arXiv cs.CL
Transformer MLP Gate Thresholds Are Couplings to a Carried Reference Direction