Decode-Time Behavioral Control in Large Language Models via Predictive Hidden-State Routing

Large language model behavior is commonly shaped through fine-tuning, RLHF, or preference optimization—methods that permanently modify model weights and entangle policy with capability. This work demonstrates that a significant portion of model behavior is instead predictable at decode time, prior to token emission, and can be selectively controlled without retraining the base model.

We introduce a decode-time behavioral control framework that learns lightweight predictors over transformer hidden states to anticipate emergent behaviors such as repetition, verbosity, and hedging before they manifest in text. These predictors enable targeted, reversible interventions during sampling, allowing behavioral steering without altering the underlying model parameters.

Key contributions include:

• Empirical evidence that multiple behavioral modes are linearly separable from intermediate hidden states, with up to 125× class separation observed for repetition.

• A multi-head decode-time control mechanism (CF-HoT) that dynamically suppresses undesirable behaviors during generation.

• A stability-focused self-improvement loop combining conservative training, multi-metric evaluation, A/B checkpoint comparison, and automatic rollback to prevent behavioral collapse.

• An open-source 8B-scale implementation demonstrating sustained quality improvements without reward hacking or loss of coherence.

This work reframes the base model as a foundational capability layer, with behavior increasingly governed by inference-time policy mechanisms rather than repeated fine-tuning. The results suggest a path toward more modular, interpretable, and reversible control of LLM behavior, with implications for efficiency, alignment, and deployment safety.

The accompanying repository provides full training code, evaluation tooling, and reproducible experiments.