Originally published on AI Tech Connect.
What you need to know A negative instruction is not a control. "Do not make things up" competes with a much stronger pull towards producing an answer, and it offers the model no alternative destination. Abstention has to be a valid output shape. The change with the most structural leverage is moving from prose to a schema in which insufficient_evidence is a legitimate terminal value, not a failure. Three different problems get called hallucination. Missing knowledge, missing context, and misread context that is present. Each has a different fix and mixing them wastes weeks. Over-refusal is a real cost. A system tuned only against fabrication becomes useless. You are choosing an operating point, not maximising one axis. Without an unanswerable eval partition you cannot measure any of this.…
Top comments (0)