OpenLongTail — component ablation

Four arms, identical backbone, identical data, identical 20 000-step budget, single seed. Each arm is evaluated exactly the way it was trained. Every delta is a paired comparison over the same 60 held-out scenes.

What each component buys

Paired mean difference per view, with 95 % bootstrap CI. The grey band is the measured noise floor, not an assumption — see below.

Absolute scores

Same five views, one small multiple each, arms in build-up order A→D.

Table view