Adaptive Repetition Controller (ARC): Behavioral Probes for Decode-Time Control and Adaptive Compute Allocation in Large Language Models

I  present the Adaptive Repetition Controller (ARC), a decode-time system for controlling behavioral artifacts and improving inference efficiency in large language models without retraining the base model.

ARC is built on the observation that several RLHF-induced behaviors—most notably semantic repetition—are predictable from hidden state representations before token emission. Using lightweight behavioral probes operating on low-dimensional projections of transformer hidden states, ARC can reliably detect these behaviors in advance and intervene during decoding.

Empirically, we demonstrate that repetition-prone states exhibit near-perfect linear separability, achieving a 125× class separation ratio, far exceeding typical probe performance reported in prior work. This indicates that RLHF induces distinct, low-dimensional policy modes that are internally represented prior to surface realization.

Beyond behavioral control, ARC repurposes these probes as efficiency oracles. Because predictable behaviors correspond to low-information, computationally simple content, the same signals can guide adaptive compute allocation, including speculative decoding, layer skipping, and early exit. In combination, these techniques yield substantial inference speedups with minimal quality degradation.

We further explore training paradigms that optimize for brevity (terse mode) and information density (dense mode) using probe-guided reward signals, and outline a pathway toward bounded recursive self-optimization through iterative probe refinement.

This repository includes the full technical paper, trained probes, example adapters, and reference implementations. ARC is intended as an open, evolving research project at the intersection of interpretability, alignment, and efficient inference.