Split Personality Training: Revealing Latent Knowledge Through Alternate Personalities

TL;DR

Split Personality Training (SPT) introduces a dual-personality approach to reveal hidden knowledge in language models by activating an honest persona that reviews responses post-generation, significantly improving detection of misbehavior.

Contribution

The paper proposes Split Personality Training, a novel fine-tuning method that creates an inactive honest persona to audit model responses, enhancing transparency and misbehavior detection.

Findings

SPT achieves 96% accuracy in detecting concealed misbehavior.

The honest persona uncovers latent biases and knowledge inaccessible to external observers.

SPT outperforms existing auditing techniques on the Anthropic benchmark.

Abstract

Detecting misalignment in large language models is challenging because models may learn to conceal misbehavior during training. Standard auditing techniques fall short: black-box methods often cannot distinguish misaligned outputs from benign ones, and mechanistic interpretability does not scale with model capabilities. We introduce Split Personality Training (SPT), which fine-tunes a second ``honest persona'' into LoRA parameters that remain inactive during normal operation. After the main model responds, we activate the LoRA adapter and insert a trigger string, enabling the honest persona to review the response while accessing the main model's latent states. We test our method on the Anthropic Auditing Game Model Organism, a benchmark where Llama-3.3-70B is trained to exploit reward hacks while concealing this behavior. SPT achieves 96% overall accuracy, whereas Anthropic reports near 0% accuracy. The honest persona reveals latent knowledge inaccessible to external observers, such as the fictional biases the compromised model was trained on.