ARC: Decode-Time Behavioral Control and Guarded Self-Optimization in Large Language Models

We present ARC (Adaptive Repetition Controller), a decode-time control framework for large language models that enables behavioral regulation, adaptive compute allocation, and bounded self-optimization without retraining the base model.

ARC is based on the empirical finding that several RLHF-induced behaviors—most notably semantic repetition—are predictable from transformer hidden states prior to token generation. Using lightweight behavioral probes operating on low-dimensional projections of hidden states, ARC can reliably detect when a model is entering low-information or policy-driven modes before those behaviors manifest in text.

We demonstrate that repetition-prone states exhibit extreme linear separability, achieving a 125× class separation ratio, far exceeding typical probing results reported in prior work. This indicates that certain alignment-style behaviors correspond to distinct, internally represented policy modes rather than diffuse stylistic tendencies.

ARC uses these behavioral predictions as a decode-time control surface, enabling targeted interventions such as repetition suppression, verbosity reduction, and policy-mode attenuation. Crucially, the same signals act as efficiency oracles, allowing the system to dynamically adjust speculative decoding, layer skipping, and early-exit decisions based on the predicted informational content of upcoming tokens.

Building on this control layer, we introduce a guarded self-optimization loop that performs conservative, short-horizon training using probe-guided reward signals. Each optimization step is instrumented with multi-metric evaluation, A/B checkpoint comparison, and automatic rollback when quality degradation is detected. This results in a stable, bounded improvement process that increases response density and efficiency without collapse or unbounded capability drift.

Importantly, ARC does not modify the underlying cognitive substrate of the base model. All adaptation occurs through decode-time control and tightly scoped optimization with explicit safeguards. The system is therefore best understood as a control-plane augmentation, not an open-ended recursive self-improving agent.

This release includes the full technical paper, trained probes, reference implementations, and example checkpoints. ARC is intended as an open research contribution at the intersection of interpretability, alignment, and efficient inference, and as a foundation for future work on inference-time control systems.