Sentence-Level Context Sensitivity as a Training-Free Detector of Unsupported Content, Evaluated Against Trained Verifiers

Name: Sentence-Level Context Sensitivity as a Training-Free Detector of Unsupported Content, Evaluated Against Trained Verifiers
Uploaded: 2026-10-06T12:27:27.804694+00:00
Duration: 368 s
Description: This research introduces Grounding-Aware Sensitivity by Perturbation (GASP), a training-free method to detect unsupported claims in Retrieval-Augmented Generation (RAG) systems. It scores each answer sentence based on how strongly its likelihood depends on the retrieved evidence, using dynamical systems theory to measure context perturbation effects.

This research introduces Grounding-Aware Sensitivity by Perturbation (GASP), a training-free method to detect unsupported claims in Retrieval-Augmented Generation (RAG) systems. It scores each answer sentence based on how strongly its likelihood depends on the retrieved evidence, using dynamical systems theory to measure context perturbation effects.

Open the full episode