{"paper":{"title":"Copy First, Translate Later: Interpreting Translation Dynamics in Multilingual Pretraining","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"In multilingual pretraining, models first copy tokens before developing general translation abilities.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Barbara Plank, Felicia K\\\"orner, Florian Eichin, Gitta Kutyniok, Maria Matveev, Michael A. Hedderich","submitted_at":"2026-04-19T22:03:29Z","abstract_excerpt":"Large language models exhibit impressive cross-lingual capabilities. However, prior work analyzes this phenomenon through isolated factors and at sparse points during training, limiting our understanding of how cross-lingual generalization emerges--particularly in the early phases of learning. To study the early trajectory of linguistic and translation capabilities, we pretrain a multilingual 1.7B model on nine diverse languages, capturing checkpoints at a much finer granularity. We use word-level translation as a testbed, introducing a novel dataset to trace how translation develops over trai"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We find that the model quickly acquires basic linguistic capabilities in parallel with token-level copying, while translation develops in two distinct phases: an initial phase dominated by copying and surface-level similarities, and a second phase in which more generalizing translation mechanisms are developed while copying is refined.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The behavioral analyses, model-component inspections, and parameter-based ablations on the chosen checkpoints and novel dataset accurately isolate translation dynamics from confounding factors such as data overlap or model architecture specifics.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Multilingual pretraining develops translation in two phases: early copying driven by surface similarities, followed by generalizing mechanisms while copying is refined.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"In multilingual pretraining, models first copy tokens before developing general translation abilities.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"d124692a45724e99173aa811e49d23a336ca79d373a5e9e95a5db13da97201ff"},"source":{"id":"2604.17633","kind":"arxiv","version":2},"verdict":{"id":"d6b52f51-0fc2-47f2-a353-e7cb68dd237d","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T05:33:30.283524Z","strongest_claim":"We find that the model quickly acquires basic linguistic capabilities in parallel with token-level copying, while translation develops in two distinct phases: an initial phase dominated by copying and surface-level similarities, and a second phase in which more generalizing translation mechanisms are developed while copying is refined.","one_line_summary":"Multilingual pretraining develops translation in two phases: early copying driven by surface similarities, followed by generalizing mechanisms while copying is refined.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The behavioral analyses, model-component inspections, and parameter-based ablations on the chosen checkpoints and novel dataset accurately isolate translation dynamics from confounding factors such as data overlap or model architecture specifics.","pith_extraction_headline":"In multilingual pretraining, models first copy tokens before developing general translation abilities."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.17633/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"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"}