arxiv
PublishedApril 18, 2026 at 4:00 AM
—neutral
Correcting Suppressed Log-Probabilities in Language Models with Post-Transformer Adapters
Publisher summary· verbatim
arXiv:2604.14174v1 Announce Type: cross Abstract: Alignment-tuned language models frequently suppress factual log-probabilities on politically sensitive topics despite retaining the knowledge in their hidden representations. We show that a 786K-parameter (approximately 0.02% of the base model) post-
Stay posted· Newsletter
A 5-min weekly brief — top movers, price watch, story of the week.
Discussion
No replies yet. Be first.
Related coverage
More from ARXIV
arxivLocal Additive Feature Attribution: A Mathematical Taxonomy and Reporting Checklist5harxivThe Steering Budget: Examples beat Knobs5harxivPolestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs5harxivRxBrain: Embodied Cognition Foundation Model with Joint Language-Visual Reasoning and Imagination5hThe Bubble Brief
WEEKLYRead language models insights every Tuesday — top movers, new releases, story of the week.
Originally published on arxiv ↗