Confidence scores play a central role in how LLM outputs are handled: they decide whether an answer is accepted, sent for review, or regenerated. But the abstract notes that scalar recalibration—the standard way of adjusting these scores—can only fix errors that are already visible through the score itself.
To address this, the authors introduce RADAR, which stands for Repair, Adopt, and Diagnose After… (the abstract cuts off there). The name suggests a broader approach than simple recalibration, but the full details are not provided in the available text.
Given the abstract, RADAR appears to be a response to a known weakness in confidence estimation. The authors argue that relying on scalar scores alone is insufficient, and they propose a method that goes beyond what recalibration can achieve.