Making AI cite its sources more precisely backfires — larger models get worse at attribution

Researchers tested whether forcing AI to cite individual sentences (instead of paragraphs or documents) improves accuracy. It doesn't. Fine-grained citations actually degrade attribution quality by 16-276% across model scales, with paragraph-level citations performing best. This means the standard assumption about citation precision has it backwards.

The field assumed tighter citations mean better verification. But the paper shows that sentence-level citations force the model to break apart the semantic dependencies it needs to properly attribute claims — it's like forcing someone to explain their reasoning one word at a time instead of one thought at a time. Larger models suffer disproportionately, suggesting that state-of-the-art systems are already optimized to reason across multiple sentences, and artificially chopping that up just breaks them. The practical implication is immediate: citation granularity is not just a UI choice, it's a fundamental constraint on model capability.