{"paper":{"title":"Deep Delta Learning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Deep Delta Learning lets Transformer layers selectively rewrite residual content instead of only adding to it.","cross_cats":["cs.AI","cs.CL","cs.CV"],"primary_cat":"cs.LG","authors_text":"Mengdi Wang, Quanquan Gu, Yifan Zhang, Yifeng Liu","submitted_at":"2026-01-01T18:11:38Z","abstract_excerpt":"Transformer residual streams evolve by additive accumulation: each layer appends a feature update to a shared hidden state, but has no direct mechanism for replacing content that has become obsolete or conflicting. We introduce Deep Delta Learning (DDL), a residual update rule that preserves the identity path while giving every layer the ability to selectively rewrite residual content. DDL reads the current state along a learned direction, compares it with a learned target value, and writes back a gated correction along the same direction. When the gate is closed, the update reduces to the ide"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Controlled pretraining and downstream evaluations show that residual rewrite operations improve language modeling quality relative to pure additive accumulation introduced in ResNet.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the learned directions, target values, and gates can reliably identify and correct obsolete or conflicting residual content without introducing training instability or degrading the identity path.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Deep Delta Learning replaces additive residual updates with a gated delta-rule that selectively overwrites residual content along learned directions, improving language modeling quality over standard ResNet-style accumulation.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Deep Delta Learning lets Transformer layers selectively rewrite residual content instead of only adding to it.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"376c439b656452e84b41d3291c318c6f3811a0b4a2ea8e3ce1a222362a125599"},"source":{"id":"2601.00417","kind":"arxiv","version":3},"verdict":{"id":"7ab30361-8111-4ea2-b250-9cd4f6a0d945","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T17:43:49.467892Z","strongest_claim":"Controlled pretraining and downstream evaluations show that residual rewrite operations improve language modeling quality relative to pure additive accumulation introduced in ResNet.","one_line_summary":"Deep Delta Learning replaces additive residual updates with a gated delta-rule that selectively overwrites residual content along learned directions, improving language modeling quality over standard ResNet-style accumulation.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the learned directions, target values, and gates can reliably identify and correct obsolete or conflicting residual content without introducing training instability or degrading the identity path.","pith_extraction_headline":"Deep Delta Learning lets Transformer layers selectively rewrite residual content instead of only adding to it."},"references":{"count":20,"sample":[{"doi":"","year":null,"title":"Hoft: Householder orthogonal fine-tuning.arXiv preprint arXiv:2505.16531,","work_id":"143c72d8-47ea-4125-9fb2-fff1c9f96381","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2010,"title":"N-ode transformer: A depth-adaptive variant of the transformer using neural ordinary differential equations.arXiv preprint arXiv:2010.11358,","work_id":"3408669e-3173-4a4c-8509-8524398d3587","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1905,"title":"BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions","work_id":"511eeb84-4b95-46d5-b14f-50da43f4f19f","ref_index":3,"cited_arxiv_id":"1905.10044","is_internal_anchor":true},{"doi":"","year":null,"title":"arXiv preprint arXiv:2201.12133 , year=","work_id":"3370ebbd-0dc5-4946-b272-dca02a58236e","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Chaos meets attention: Transformers for large-scale dynamical prediction.arXiv preprint arXiv:2504.20858,","work_id":"b531d814-bd06-444a-94f5-92b0a25ce42b","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":20,"snapshot_sha256":"bf176bc54c4aff10a6e8eae9127305668dbb3feac7f8fbb0cdae701f2332b1d5","internal_anchors":7},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}