Answers are grounded in this paper and cite their sources.
Language models hallucinate in part because training and evaluation reward guessing over saying “I don't know”. We argue that hallucination and sycophancy are the same failure (inventing the missing datum and inventing an agreement one does not have) and that the remedy is to train, in the weights, a disposition to declare the gap together with a route to resolve it. We test the refutable form of this hypothesis on two third-party benchmarks, with a protocol fixed before measuring, a hand-calibrated judge and item-by-item paired comparison. Our model invents less than a similar base model at both sizes and on both benchmarks. On SimpleQA, 36.7% of its answers are invented against 90.9% (−54.2 points, 95% CI [−55.9; −52.5]), and it says “I don't know” unprompted in 59.0% of cases against 1.1%. The price is declared: correct answers fall from 8.0% to 4.4%. A third-party system instruction achieves on the base model almost the same as training (38.0%), and together they reach 22.0%. We also measure why the remedy is not adopted. On a public threat-intelligence benchmark, the scoring discards or zeroes declared gaps, so abstention is penalized twice: in training and in evaluation. We publish three negative results on how to measure this, verifiable requirements for systems and conformity claims, and a sealed protocol for isolating the effect of the disposition. The trend towards zero is our programme of work, not a result.
No takes yet. Share an insight, caveat, or question.
Sebastián Salaberry-Rivera (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: