Inferences are both commonplace and elusive in reading. The key questi
on is: What principles separate the commonplace from the elusive? Alth
ough those who argue for on-line goal-directed inferences may be corre
ct, the evidence in favor of such inferences is yet not substantial an
d is open to other interpretations. Higher level inferences may be res
tricted in part because, unlike syntactic processes, they do not opera
te in response to simple memory symbols, but depend on complex composi
tional representations that might not always be available.