What does this research mean for the field?

HeartScale Coherence Rejection Sampling (HCRS) enables inference-time alignment of autoregressive language models without requiring learned reward models, human preference data, or post-training interventions. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

The research aims to introduce HeartScale Coherence Rejection Sampling as a method for ensuring token alignment in language models during inference.

June 3, 2026Open Access

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

Key Points

The research aims to introduce HeartScale Coherence Rejection Sampling as a method for ensuring token alignment in language models during inference.
Developed a per-token acceptance filter that maintains alignment constraints without external interventions.
Utilized a criterion based on the weighted harmonic mean of vortex drift and entropy excess.
Implemented a systematic progressive-temperature resampling for rejected tokens.
Achieved effective alignment of tokens during inference without the need for human preference data.
Continuous monitoring demonstrated robustness across various training runs.

Abstract

We present HeartScale Coherence Rejection Sampling (HCRS), a per-token acceptance filter for autoregressive language models that enforces alignment constraints at inference time without a learned reward model, human preference data, or any post-training intervention. The acceptance criterion is the weighted harmonic mean of vortex drift and entropy excess divided by a session-level capacity scalar, evaluated against the Ma'at balance predicate. Rejected tokens trigger progressive-temperature resampling. Production-implemented and continuously monitored across training runs via the HeartScaleViolationsSuite.

Read Full Paperexternally

AI से पूछें

Bookmark

View Full Paper