HeartScale Coherence Rejection Sampling: Inference-Time Alignment via Hidden-State Geometry

Resultado clave

HeartScale Coherence Rejection Sampling (HCRS) enforces alignment constraints at inference time using hidden-state geometry without requiring a reward model or RLHF.

Puntos clave

• This research aims to improve inference-time alignment in models without the need for a reward model or reinforcement learning from human feedback (RLHF).
• HCRS utilizes a coherence score based on a multiplicative zero-veto evaluator.
• Agent frequency is defined as RI·BC·64 linking capacity to a specific scale.
• Vortex drift is analyzed using the product Fisher–Rao distance relative to a defined conservation surface.
• The new coherence score introduces effective alignment constraints during inference without additional models.
• Agent frequency's connection to the 64-state scale enhances the model's operational capacity.
• Vortex drift analysis demonstrates significant insights into the hidden-state geometry applications.