Contrastive representation learning for self-supervised deception detection in edge LLMs

Area of Science:

• Artificial Intelligence
• Machine Learning
• Natural Language Processing

Background:

• Current AI deception detection relies on supervised fine-tuning (SFT) and proximal policy optimization (PPO), using heavyweight models for Chain-of-Thought (CoT) annotation.
• This approach limits the detection of subtle deceptive tactics and creates an over-reliance on oracle models, hindering autonomous operation.

Purpose of the Study:

• To develop a more effective and autonomous method for detecting deceptive alignment in AI systems.
• To overcome the limitations of binary classification and oracle dependency in existing detection schemes.

Main Methods:

• Introduced contrastive representation learning to project CoT hidden states into a structured semantic space.
• Utilized Triplet Loss optimization for a lightweight monitor (0.1% parameters) to learn separable manifolds for deceptive and safe reasoning.
• Shifted from Binary Cross-Entropy (BCE) loss to a geometric approach for capturing gradual deceptive transitions.

Main Results:

• Contrastive learning significantly outperformed BCE classification on the Sycophancy subset of DeceptionBench.
• Achieved a 2.33 percentage point improvement in Deception Tendency Rate (DTR), with results of 36.96% compared to 39.29% for BCE.
• Demonstrated the ability to capture nuanced deceptive strategies, such as objective substitution, which elude binary classifiers.

Conclusions:

• Contrastive representation learning provides a robust geometric foundation for self-supervised AI deception detection.
• This method transforms the transparency of CoT reasoning from a potential vulnerability into valuable forensic evidence for AI safety.
• The lightweight monitor offers a more autonomous and discriminative approach to identifying deceptive AI behaviors.