Same time, same object, different value: probing a small language model for conflicts inside the context window
A pre-registered pilot on Qwen2.5-0.5B. A first protocol met all its criteria (AUROC 1.00), but post-hoc controls showed the result was lexical. A second protocol, trained on colours and materials and tested on sizes, an attribute never seen in training, met all its criteria again: AUROC 0.99 pooled, 0.98 against tempo...