Transformer MLP Gate Thresholds Are Couplings to a Carried Reference Direction

The corpus-mean direction of a transformer’s residual stream is usually treated as a nuisance andsubtracted before analysis. We present evidence that it is a functional component: the reference againstwhich the MLP gate population sets its operating point. In Phi-2, an exact decomposition of restinggate pre-activations shows that the resting inhibition of >99.9% of gates is carried by the couplingw · bL to the carried mean direction, at 50–60× the explicit bias parameter, and this resting termrank-predicts each neuron’s firing rate at ρ ≈ 0.95. Causally, removing the stream’s projection on thereference multiplies above-threshold firing by ∼9×, dose-monotonically, at 30–40× a norm-matchedcontrol; a random-initialised twin is flat, and replacement tests show the function is carried by thedirection, not the magnitude or token content. The decomposition replicates on four further familiesspanning both gate types (GELU with an explicit gate bias, bias-free SwiGLU), and the dose-responseon two of them. Across an eight-model scan the mechanism is present in every GELU and SiLUfamily and absent only in OPT, where an opposing LayerNorm bias cancels the carried reference.Gate thresholds are implemented as couplings to a constant the network itself builds and maintains,not as bias parameters.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-07-22
DOI
https://doi.org/10.5281/zenodo.21498412
Primary Topic
Neuroscience and Neuropharmacology Research
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
preprint

Transformer MLP Gate Thresholds Are Couplings to a Carried Reference Direction

Olli Tuomi
Zenodo (CERN European Organization for Nuclear Research)
Neuroscience and Neuropharmacology Research
preprint

Transformer MLP Gate Thresholds Are Couplings to a Carried Reference Direction

Olli Tuomi
preprint en

Abstract

The corpus-mean direction of a transformer’s residual stream is usually treated as a nuisance andsubtracted before analysis. We present evidence that it is a functional component: the reference againstwhich the MLP gate population sets its operating point. In Phi-2, an exact decomposition of restinggate pre-activations shows that the resting inhibition of >99.9% of gates is carried by the couplingw · bL to the carried mean direction, at 50–60× the explicit bias parameter, and this resting termrank-predicts each neuron’s firing rate at ρ ≈ 0.95. Causally, removing the stream’s projection on thereference multiplies above-threshold firing by ∼9×, dose-monotonically, at 30–40× a norm-matchedcontrol; a random-initialised twin is flat, and replacement tests show the function is carried by thedirection, not the magnitude or token content. The decomposition replicates on four further familiesspanning both gate types (GELU with an explicit gate bias, bias-free SwiGLU), and the dose-responseon two of them. Across an eight-model scan the mechanism is present in every GELU and SiLUfamily and absent only in OPT, where an opposing LayerNorm bias cancels the carried reference.Gate thresholds are implemented as couplings to a constant the network itself builds and maintains,not as bias parameters.

Zenodo (CERN European Organization for Nuclear Research)
Scientific Solutions (United States) (US)
Neuroscience and Neuropharmacology Research
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.