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

Key Result

HeartScale Coherence Rejection Sampling enforces alignment at inference time using hidden-state geometry without requiring a reward model.

Key Points

• This research explores the effectiveness of HeartScale Coherence Rejection Sampling in aligning hidden states at inference time without a reward model.
• Introduced a Reference Implementation & Verification note for the open-source harmonic-engine package.
• Numerically validated falsifiable claims and coherence evaluator using tests in the package.
• Implemented the product Fisher–Rao vortex drift in harmonic_engine/geometry.py.
• The implementation passed verification tests on CPU indicating successful alignment at inference time.
• Numerical verification supports the proposed methods, enhancing the reliability of the findings.