Tag
This paper investigates how toxic lexical perturbations in prompts reduce the factual accuracy and increase uncertainty of LLMs, and uses attribution-graph analyses to trace internal changes. It finds that increasing toxicity amplifies perturbation-sensitive variant nodes while core reasoning nodes remain invariant.
This paper investigates why Retrieval-Augmented Generation (RAG) systems fail despite having access to correct evidence. Using circuit tracing and attribution graphs, the authors find that correct predictions exhibit deeper reasoning paths and more distributed evidence flow, while failures show shallow and fragmented patterns. They propose a graph-based error detection framework and targeted interventions to improve RAG reliability.