RESEARCHInvestigateNEXT 12 MONTHS
The Invisible Editorial Layer: Formalizing Undisclosed Inference-Time Steering, Probability Placement, and the Attribution Problem in Deployed Language Models
arXiv cs.CL — Computation and Language
Factual evidence
What the source reports
ArXiv paper shows how undisclosed inference-time steering masks underlying model behavior, complicating auditability and evaluation.
Open source