On CATT Rule Routings (RR), the strongest selected CATT gradient baseline recovers 22.35% of the relevant documents at Recall@10, while BM25 recovers 99.80%. A
Can final hidden states recover the training documents that teach a language model a rule? CATT's weak representation baseline leaves open whether useful inform
Retraining-based evaluation asks how removing training documents changes a model's outputs. We measure what such removal experiments produce on CATT Rule Routin
Low-rank projected gradient attribution on CATT performs strongly on direct constant-binding memorization (`Function Bindings with distractors`, `FB-distr`) yet
When wording changes make lexical source retrieval ineffective, can a fixed contextual readout still recover the relevant rule documents? On CATT Rule Routings
Can a compact model readout preserve strong rule-support retrieval across CATT's Rule Routings and Function Bindings with distractors? We combine established ID
Recovering a training source does not establish that a model can use its rule. We test this distinction on CATT Rule Routings after supplied attribute substitut
On CATT Rule Routings, model-based scores recover the documents that express a query's rule, and gradient scores are sensitive to the queried answer. We ask wha