{"paper":{"title":"Training Infinitely Deep and Wide Transformers","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Gradient flow in the conditional Wasserstein metric converges to global minima for infinitely deep and wide transformers when the initial loss is small enough and the attention NTK is injective.","cross_cats":["cs.AI","cs.LG","stat.ML"],"primary_cat":"math.OC","authors_text":"Gabriel Peyr\\'e, Maarten V. de Hoop, Rapha\\\"el Barboni, Takashi Furuya","submitted_at":"2026-05-17T21:30:13Z","abstract_excerpt":"Transformers have become the dominant architecture in modern machine learning, yet the theoretical understanding of their training dynamics remains limited. This paper develops a rigorous mathematical framework for analyzing gradient-based training of transformers in the mean-field regime, where both the depth (number of layers) and width (number of attention heads) tend to infinity. While ResNet training can be understood as controlling a neural ODE, transformer training corresponds to controlling a neural PDE, due to the coupling of multiple token distributions through the attention mechanis"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Under the NTK injectivity assumption (linear independence of log-sum-exp functions modulo affine functions), gradient flow in the conditional Wasserstein metric converges to global minima whenever the initial loss is sufficiently small.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The mean-field limit and the NTK injectivity condition remain valid approximations for the finite but large transformers that are actually trained; this premise enters when the convergence theorem is stated after the injectivity characterization.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Develops a mean-field neural PDE model for transformer training, proves forward-pass well-posedness via function-space ODEs, derives conditional Wasserstein gradients, and shows global convergence of gradient flow under an NTK injectivity condition equivalent to linear independence of log-sum-exp fu","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Gradient flow in the conditional Wasserstein metric converges to global minima for infinitely deep and wide transformers when the initial loss is small enough and the attention NTK is injective.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"0ef8e9c083ae82fb032841b05de678609a723a0bae106e4075ad4349d14ce7f2"},"source":{"id":"2605.17660","kind":"arxiv","version":1},"verdict":{"id":"3204ae40-3192-4bd1-b78a-c06e6933c0b5","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T22:08:22.341277Z","strongest_claim":"Under the NTK injectivity assumption (linear independence of log-sum-exp functions modulo affine functions), gradient flow in the conditional Wasserstein metric converges to global minima whenever the initial loss is sufficiently small.","one_line_summary":"Develops a mean-field neural PDE model for transformer training, proves forward-pass well-posedness via function-space ODEs, derives conditional Wasserstein gradients, and shows global convergence of gradient flow under an NTK injectivity condition equivalent to linear independence of log-sum-exp fu","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The mean-field limit and the NTK injectivity condition remain valid approximations for the finite but large transformers that are actually trained; this premise enters when the convergence theorem is stated after the injectivity characterization.","pith_extraction_headline":"Gradient flow in the conditional Wasserstein metric converges to global minima for infinitely deep and wide transformers when the initial loss is small enough and the attention NTK is injective."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.17660/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T22:31:19.483169Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"cited_work_retraction","ran_at":"2026-05-19T22:22:11.642152Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T22:20:57.559089Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"citation_quote_validity","ran_at":"2026-05-19T21:49:44.270375Z","status":"skipped","version":"0.1.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T21:33:23.541285Z","status":"skipped","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T21:21:57.458298Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"bac2ec79408c39f98fce8a078365c1009a4553dbe1fa1e5d7a30ef8a506dd882"},"references":{"count":53,"sample":[{"doi":"","year":2023,"title":"Transformers learn to imple- ment preconditioned gradient descent for in-context learning","work_id":"42670431-9028-4b66-889a-652920100a66","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"What learning algorithm is in-context learning? investigations with linear models","work_id":"f67e4fac-adc7-43d4-b334-b8b9d73e40ac","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"A convergence theory for deep learning via over-parameterization","work_id":"00507d70-4223-4a3a-9734-7857dbf05620","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2008,"title":"Transport equation and Cauchy problem for non-smooth vector fields","work_id":"89b697c2-19c1-4aea-9f8b-ee3e65fac016","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2062,"title":"A user’s guide to optimal transport","work_id":"68f196ac-7a64-491a-a457-3a7adbbd7ede","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":53,"snapshot_sha256":"c623850a99f2bf16a3b1bd791d3031a39a3fc798de23afad680d4b9fdfef9187","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"9f10c5a324cfd2b25af36ad09efb6164574e418af8cd7ff37bf1f8d1b6d2aee2"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}