Calibrating Verbalized Confidence with Self-Generated Distractors

Researchers introduce DINCO (Distractor-Normalized Coherence), a method to improve confidence calibration in large language models by using self-generated alternative claims to reduce overconfidence bias. The approach addresses LLM suggestibility issues that cause models to express high confidence on low-accuracy outputs, potentially improving AI safety and trustworthiness.

• →Large language models suffer from miscalibrated confidence, showing high confidence on incorrect outputs due to suggestibility bias.
• →DINCO method uses self-generated distractors to normalize confidence estimates and reduce overconfidence in AI systems.
• →The approach combines generator-validator disagreement with consistency-based estimates to improve calibration accuracy.
• →DINCO outperforms traditional self-consistency methods while requiring significantly fewer inference calls (10 vs 100).
• →Better confidence calibration is crucial for building trust in AI systems and improving their real-world deployment safety.