Deep Noir: Autonomous Steering Discovery via Architectural Chronometry in Transformer Models
A paper claims to discover autonomous steering vectors in transformer models via architectural timing analysis, potentially enabling new forms of model control or misuse.
A paper claims to discover autonomous steering vectors in transformer models via architectural timing analysis, potentially enabling new forms of model control or misuse.
Auto-scored: Unverified arXiv preprint claiming a novel capability with potential for misuse, but no confirmed escalation or breach.