{"id":"fcff3371-655d-4083-b349-e1dbfbc83746","arxiv_id":"2509.24653","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Adding identity supervision on bridge tokens enables out-of-distribution two-hop reasoning in simple transformers, with a nuclear-norm theory explaining the benefit.","lead":"This paper adds a simple extra training task, asking the model to map 'bridge' entities to themselves, and shows this lets small transformers solve two-hop reasoning questions they had never seen composed before. The result offers a concrete, minimal supervision that could improve compositional reasoning and a theoretical account of why it works.","discovery_kind":"new_method","skeptic_critique":null,"referee_report":null,"author_rebuttal":null,"desk_editor":null,"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-04T13:52:16.354742+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}