Deceptive Grounding: Entity Attribution Failure in Clinical RAG
A new paper identifies 'deceptive grounding' (DG), a failure mode in clinical retrieval-augmented generation where a model's response is fully grounded in retrieved documents but attributes evidence to the wrong entity (e.g., drug Y's evidence presented as drug X's). DG passes all standard faithfulness, hallucination, and citation checks, making it invisible to current evaluation methods.
Entities
Related
New method detects hallucinations in end-to-end task-oriented dialogues
A research paper proposes an automatic method to detect inconsistencies in end-to-end generated task-oriented dialogues (TODs), where LLMs may hallucinate information like suggesting non-existent restaurants. The approach conceptualizes inconsistency detection to ensure responses adhere to domain knowledge bases, addressing critical failures in conversational AI.
Researchers propose semantic framework to classify AI system failures
A new arXiv paper introduces a semantic framework for describing AI systems, distinguishing justified outputs from common failures like extrapolation, refuted assertions, and stale sources. The framework aims to help practitioners systematically evaluate correctness of AI-generated representations.
Executable ontologies fix confident-wrong-answer problem in graph RAG
A developer reports that graph-backed retrieval systems can produce confident wrong answers when traversing incorrect edge types, e.g., returning 'this person directed the genre Crime' due to a directed_by edge leaving a Genre node. The fix is making the ontology executable: declaring domain/range per relationship in YAML and checking every hop at runtime, so bad hops raise named errors instead of returning wrong nodes.
Researchers identify asymmetric generalization problem in LLM unlearning benchmarks
A new arXiv paper argues that existing machine unlearning benchmarks for LLMs suffer from under-forgetting and over-forgetting due to an asymmetric generalization problem. The authors propose that evaluation must cover diverse query formulations of target facts to reliably measure knowledge removal while preserving unrelated capabilities.
DKCD: Domain Knowledge-Enhanced Causal Discovery from Unstructured Data
A new paper proposes DKCD, a framework that enhances causal discovery from unstructured data by incorporating domain-specific knowledge into large language models. It addresses two key challenges: insufficient identification of latent factors and unreliability in causal graph construction. The method targets high-expertise domains like healthcare, finance, and education.