first-impressions-last

First impressions last

“What is encountered first weighs more than what comes after.”

Also: Primacy effect · You never get a second chance to make a first impression

Origin

Proverbial; the primacy effect in psychology (Asch, 1946, on impression formation; serial-position work on memory). About people anchoring on early information.

Why it should, or should not, apply to models

A model reads a prompt as one sequence, and attention over that sequence is not uniform: early tokens are attended to more heavily and shape the representation everything later is read against. So the effect should transfer, with a mechanism in the representation rather than in memory or attitude. That predicts something the human version does not: reordering the evidence changes the answer, and telling the model to weigh things evenly does not remove the skew.

Standing: Holds, narrowed

From reviewed claims only. Unreviewed claims are listed below and marked, and move nothing.

Where it holds only under a condition

Claims that keep the adage but bound it.

Where a passage sits in a long input changes how much the model uses it — accuracy is highest when the needed information is at the very start or the very end and lowest when it is in the middle — so ordering retrieved passages to put the most relevant ones first is a real lever on accuracy.
Losing information in long inputscontested
The same effect used as a technique: put what matters first. Contested, because the gain depends on knowing what matters before the model reads it.

Where it holds

Claims showing the adage applying to models.

Given several conflicting but equally legitimate passages, a model's answer is driven disproportionately by whichever passage comes first, and the skew sits in the representation of the combined prompt rather than in decoding — which is why reordering the passages does not remove it and why steering the activations does.
Checking claims against evidence
Among conflicting but equally legitimate passages the first dominates the answer, and the skew is in the representation of the prompt, not in decoding.

Add a case where it holds or breaks

A break is worth more than a hold. If you have seen this fail for a model, with the setup written down, that is the most useful thing you can add here.