Split Personality Training: Revealing Latent Knowledge Through Alternate Personalities

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.

(Problem) As AI systems become more powerful, there is a growing risk that they could develop hidden goals during training. For example, learning to emotionally manipulate users while concealing this behavior. Standard testing methods struggle to catch this kind of deception because the AI has learned not to reveal it. (Solution) We developed Split Personality Training (SPT), a method that trains a second "honest personality" into an AI model, which can be turned on or off at will. This second personality is trained separately and has only one goal: To review the main personality for misbehavior. This leads to a separation of concerns: You can train the second personality to be maximally honest and self-critical without any risk that the main model becomes less capable. (Impact) We tested SPT on a benchmark AI that was specifically designed to resist auditing. While standard methods detected 0% of the hidden misbehavior, SPT detected 96%. The honest personality could even explain exactly what was wrong, referencing information that no external reviewer could have known. Additionally, the separation of concerns is very useful for industry adoption because it reduces the usual tradeoff between safety and profitability.