CF-HoT: Decode-Time Behavioral Control for Language Models via Per-Token Risk Prediction

This repository contains the complete, corrected technical reference and validated implementation of CF-HoT (Control Field Theory of Hot Tokens), a decode-time behavioral control system for large language models.

CF-HoT operates by training lightweight risk-prediction heads on transformer hidden states to anticipate undesirable behaviors—such as repetition, hedging, verbosity, and sycophancy—before they occur. These predictions are used to intervene directly on token logits during generation, enabling real-time behavioral steering without modifying the base model’s weights.

The core validated result demonstrates 125× separation in predicting imminent repetition using a per-token labeling methodology and fiber projections across all transformer layers. This work corrects architectural and methodological errors present in earlier documentation, including the use of dense per-token supervision, multi-layer aggregation, and decode-time logit intervention rather than attention modification.

The archive includes:

• A complete, corrected technical reference

• Fully specified training and inference code

• Reproducible experiments and separation metrics

• Multi-head architecture extending the approach to additional behavioral dimensions (with validation status clearly documented)

• Integration with bounded recursive self-improvement (RSI) and tokenization co-evolution (Loop 4)

CF-HoT demonstrates that behavioral control in language models can be achieved at inference time, preserving model capabilities while suppressing degenerate behaviors. The methodology is fully reproducible for all validated components and is intended as an engineering reference for researchers exploring inference-time control, model interpretability, and safe self-improvement.