Researchers presented the Inoculation Midtraining method for controlling models

Researchers proposed the Inoculation Midtraining technique, designed to limit undesirable behavior of neural networks. The method is based on introducing a special neologism-token into the training process, which binds potentially hazardous or incorrect responses to a strictly defined context.

During experiments, the authors found that using the neologism allows reducing the level of model disorientation while preserving its basic characteristics, such as speech style and general cognitive abilities. The technique is aimed at localizing "incorrect" behavior within the model, preventing its spread to other user interaction scenarios.

Testing results showed that the current implementation of the method is less effective than the standard Inoculation Prompting approach. Researchers note the algorithm's high sensitivity to hyperparameters, which requires additional tuning to achieve stable safety metrics in various architectures.

This approach represents an attempt to solve the model "alignment" problem at the stage of their fine-tuning. Unlike external filters, the method tries to integrate control mechanisms directly into the model's weights, using linguistic markers to control the activation of certain behavior patterns.