Closed Gates Leak: Decomposing Adjacent-Layer Cancellation in Gated Language Models
Source check: Source checked. Not independent scientific confirmation.
Finding: Across ten gated models closed-gate neurons carry most of the average damping; annealing their negative activation side removes damping/cancellation in tested continued training. Sixteen language models are analyzed, including ten gated models; counterfactual and continued-training controls test mechanism, with limited small-model interventions.
Technology: Controlled decomposition of activations in gated transformer MLPs attributes adjacent-layer damping to closed-gate neurons; gate-annealing interventions test the mechanism.
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- Controlled decomposition of activations in gated transformer MLPs attributes adjacent-layer damping to closed-gate neurons; gate-annealing interventions test the mechanism.
- Test and result
- Sixteen language models are analyzed, including ten gated models; counterfactual and continued-training controls test mechanism, with limited small-model interventions.
- Limitation / still unknown
- The reviewed source does not establish prospective benefit, broad external validity, or replication beyond the reported study.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/html/39068v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review date: 22 Nov 2026. Review scope: Primary result-bearing material for the stated method, result and evaluation, with bibliographic version information..
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.