Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces a supervised learning framework for LLM uncertainty quantification in content moderation, training a meta-model on LLM Performance Predictors (LPPs) derived from LLM outputs like log-probabilities and entropy. The framework enables cost-aware selective classification, escalating high-risk cases for human review while automating others. Experiments across various LLMs (Gemini, GPT, Llama, Qwen) on multimodal and multilingual moderation tasks demonstrate improved accuracy-cost trade-offs compared to existing uncertainty estimators.
Stop blindly trusting LLMs for content moderation: a meta-model using LLM Performance Predictors slashes costs by knowing when to escalate to human review.
As LLMs are increasingly integrated into human-in-the-loop content moderation systems, a central challenge is deciding when their outputs can be trusted versus when escalation for human review is preferable. We propose a novel framework for supervised LLM uncertainty quantification, learning a dedicated meta-model based on LLM Performance Predictors (LPPs) derived from LLM outputs: log-probabilities, entropy, and novel uncertainty attribution indicators. We demonstrate that our method enables cost-aware selective classification in real-world human-AI workflows: escalating high-risk cases while automating the rest. Experiments across state-of-the-art LLMs, including both off-the-shelf (Gemini, GPT) and open-source (Llama, Qwen), on multimodal and multilingual moderation tasks, show significant improvements over existing uncertainty estimators in accuracy-cost trade-offs. Beyond uncertainty estimation, the LPPs enhance explainability by providing new insights into failure conditions (e.g., ambiguous content vs. under-specified policy). This work establishes a principled framework for uncertainty-aware, scalable, and responsible human-AI moderation workflows.