A new unsupervised method called OAT tackles failure attribution in LLM-based agentic systems by training only on successful trajectories and using one-class learning with neural controlled differential…

#LLMAgents #PromptEngineering #AIDebugging #GenerativeAI
https://arxiv.org/abs/2607.12747

A position paper critiques faithfulness metrics for feature-importance XAI, arguing structural misalignments lead to unfaithful explanations and overreliance on incorrect predictions. Authors propose rethinking XAI…

#XAI #ExplainableAI #MachineLearning #AIDebugging
https://arxiv.org/abs/2610.12493