Knowing When to Escalate: Disagreement Patterns for Uncertainty-Aware Content Moderation
What it is
This paper proposes Reasoning Trace Divergence Analysis (RTDA), a method that embeds model-generated justifications from LLM agents to classify disagreement patterns in content moderation. It distinguishes between value-laden conflicts and simple reliability issues by analyzing the similarity of reasoning traces rather than just output distributions, enabling better prediction of human annotator disagreement.
The problem it solves
Standard uncertainty methods treat inter-agent disagreement as a reliability problem to be removed, which is inappropriate for value-laden tasks where disagreement reflects irreducible value pluralism, leading to poor escalation decisions.
What you could use it for
- Identifying value-laden content moderation cases that require human review
- Improving the reliability of agentic LLM moderation systems by filtering out irreducible value pluralism
- Post-processing reasoning traces to condition human escalation on disagreement structure
Why it matters
If effective, RTDA could significantly reduce unnecessary human escalations in content moderation by accurately distinguishing between cases where agents disagree due to value conflicts versus those where they disagree due to model uncertainty or error.
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.