IC-1630Llama and Pythia models represent entity-attribute bindings via additive binding id vectors that form a continuous subspace with metric structure
Using causal mediation analysis, the paper shows that Llama and Pythia models solve binding tasks by attaching abstract binding id vectors to entity and attribute token activations. The binding function is additive (entity/attribute representation plus binding id vector), the mechanism is factorizable (swapping one entity's activation changes only that entity's binding), and position-independent (permuting token positions does not change bindings). Furthermore, binding id vectors occupy a continuous subspace: linear combinations of valid binding ids are themselves valid, and the distance between two binding id vectors determines the model's ability to distinguish them, with nearby vectors causing near-chance accuracy and distant vectors yielding high accuracy.
Evidence
interventional
Key metric
Table 1 (Llama-13B, n=2 capitals): control 0.99/1.00, attribute intervention 0.00/0.03, entity intervention 0.00/0.01, both 0.97/1.00 (mean median-calibrated accuracy for e0/e1). Table 2 (Llama-30B, n=3 capitals, transfer): capitals 0.88, parallel 0.87, shapes 0.71, fruits 0.80, bios 0.47, zeros 0.30, random 0.31.
Caveat
The mechanism is not universal: an MCQ task uses an alternative direct binding mechanism. The bios task has multi-token attributes, so attribute factorizability does not hold there. Llama-65B was excluded from some experiments for computational reasons.