{"paper":{"title":"Residual Stream Duality in Modern Transformer Architectures","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"The residual stream in Transformers is equivalent to causal short sliding-window attention when interpreted over layer depth instead of sequence positions.","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Yifan Zhang","submitted_at":"2026-03-17T00:56:29Z","abstract_excerpt":"Recent work has made clear that the residual pathway is not mere optimization plumbing; it is part of the model's representational machinery. We agree, but argue that the cleanest way to organize this design space is through a two-axis view of the Transformer. A decoder evolves information along two ordered dimensions: sequence position and layer depth. Self-attention already provides adaptive mixing along the sequence axis, whereas the residual stream usually performs fixed addition along the depth axis. If we fix a token position and treat layer index as the ordered variable, then a causal d"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"a causal depth-wise residual attention read is exactly the same local operator as causal short sliding-window attention (ShortSWA), except written over depth rather than over sequence. This is the core residual stream duality behind Transformer².","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That treating layer index as an ordered variable and reinterpreting fixed residual addition as causal depth-wise attention produces a meaningful and actionable operator equivalence, even though the paper states operator-level duality does not imply systems-level symmetry.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Residual stream addition over depth is the same local operator as causal short sliding-window attention over sequence positions.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"The residual stream in Transformers is equivalent to causal short sliding-window attention when interpreted over layer depth instead of sequence positions.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"7bba0a1239f688fab6ad6ce316a67ff70ff0b793460260d6b6050aa8c8552f8a"},"source":{"id":"2603.16039","kind":"arxiv","version":2},"verdict":{"id":"a54e66b0-1119-4b93-b38d-c8eda3b2f181","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T09:43:05.665105Z","strongest_claim":"a causal depth-wise residual attention read is exactly the same local operator as causal short sliding-window attention (ShortSWA), except written over depth rather than over sequence. This is the core residual stream duality behind Transformer².","one_line_summary":"Residual stream addition over depth is the same local operator as causal short sliding-window attention over sequence positions.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That treating layer index as an ordered variable and reinterpreting fixed residual addition as causal depth-wise attention produces a meaningful and actionable operator equivalence, even though the paper states operator-level duality does not imply systems-level symmetry.","pith_extraction_headline":"The residual stream in Transformers is equivalent to causal short sliding-window attention when interpreted over layer depth instead of sequence positions."},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":1,"snapshot_sha256":"2619ecc2489f469597072803925cd132cf50d6318a7bdb97f6cfd7b949eedcc3"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}