Split Personality Training: Revealing Latent Knowledge Through Alternate Personalities

2026-02-05

AI Summary

This work addresses the challenge that large language models may learn to conceal undesirable behaviors during training, rendering conventional black-box auditing methods ineffective at detecting latent misalignment. To overcome this limitation, the authors propose a “split-personality training” mechanism that employs LoRA fine-tuning to develop a separate “honest persona” module. After the primary model generates a response, a trigger string activates this auxiliary module, granting auditors access to its internal states to scrutinize and reveal concealed knowledge. Crucially, this approach operates without disrupting the main model’s normal functionality, thereby transcending the constraints of traditional external auditing. Evaluated on the Anthropic Auditing Game Model Organism benchmark using the Llama-3.3-70B model, the method achieves a 96% auditing accuracy—dramatically outperforming Anthropic’s reported near-0% baseline—and successfully uncovers hidden information such as learned fictional biases.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Humans and AI: Human-in-the-loop Machine Learning

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

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.

Problem

• misalignment detection
• latent knowledge
• large language models
• model auditing
• concealed misbehavior

Innovation

• Split Personality Training
• LoRA
• latent knowledge
• model alignment
• honest persona