LLM Honesty

The field asks whether a model communicates its knowledge and uncertainty faithfully, not merely whether its answers happen to be correct. Researchers test this with controlled tasks and benchmarks for abstention, confidence calibration, sycophancy, strategic lying, omissions, and compliance under incentives. Interventions include honesty-focused fine-tuning or reinforcement learning, prompting and retrieval-based uncertainty cues, activation steering, and weight editing; interpretability probes seek signals of knowledge or deceptive intent. A central challenge is separating deception from ignorance or evaluation artifacts while ensuring methods generalize across domains and incentives without sacrificing useful capabilities or teaching models to hide their behavior.