IC-047Tulu-2-13B's entity-attribute binding partially relies on token order as a shortcut, degrading in nested orderings where order and semantic binding conflict
Across five template orderings (series, cross, reverse, coref, nested), propositional probes perform well in all except nested (ab21), where EM accuracy drops to 44% (vs. 25% for random binding and 0% for pure order-based binding). The paper states the model 'relies on shortcuts such as order information to capture binding.' On a second nested dataset (nested_2), prompting fails catastrophically while probes maintain similar performance. Constraining propositions to unique entities improves nested accuracy to 76%. The binding subspace is not sensitive to distance between tokens but is affected by order.
Evidence
correlational
Key metric
Nested ordering EM accuracy 44% (probes) vs 25% (random) vs 0% (pure order); 76% with unique-entity constraint; prompting fails catastrophically on nested_2
Caveat
The 44% is interpreted as a mixture of order and semantic binding; the authors note the 50-dim subspace may contain spurious directions contributing to the order sensitivity.