{"paper":{"title":"Make It Long, Keep It Fast: End-to-End 10K Long User Behavior Sequence Modeling for Billion-Scale Douyin Recommendation","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"Stacked cross-attention and shared batching let recommendation models process 10,000-length user histories at production speed.","cross_cats":["cs.IR"],"primary_cat":"cs.LG","authors_text":"Beichuan Zhang, Bo Sun, Feng Zhang, Hangyu Wang, Jia-Qi Yang, Jinan Ni, Lin Guan, Qiwei Chen, Xiaowen Li, Xiao Yang, Xuanyuan Luo, Yi Cheng, Yuhang Qi, Zhifang Fan, Zhishan Zhao","submitted_at":"2025-11-08T17:22:54Z","abstract_excerpt":"Short-video recommenders such as Douyin must exploit extremely long user behavior histories without breaking latency or cost budgets. We present an end-to-end industrial recommender system that scales long-sequence recommendation modeling to 10K-length histories in production. First, we introduce Stacked Target-to-History Cross Attention (STCA), which replaces history self-attention with stacked cross-attention from the target to the history, reducing complexity from quadratic to linear in sequence length and enabling efficient end-to-end training over long user behavior sequences. Second, we "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Deployed at full traffic on Douyin, our system delivers significant improvements on key engagement metrics while meeting production latency, demonstrating a practical path to scaling end-to-end long-sequence recommendation to the 10k regime.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the length-extrapolative training strategy (train on shorter windows, infer on 10k) generalizes without performance loss and that stacked target-to-history cross attention captures the necessary preference signals from long histories.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Douyin deploys stacked target-to-history cross attention and request-level batching to scale end-to-end recommendation modeling to 10k-length histories, observing scaling-law gains and live engagement improvements.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Stacked cross-attention and shared batching let recommendation models process 10,000-length user histories at production speed.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"6c9fe21bb1b197fced8f030413e920a1b111de53a1a18b417aeedf10f682fe5d"},"source":{"id":"2511.06077","kind":"arxiv","version":3},"verdict":{"id":"0d78b2d3-d76d-4426-bce8-7c5512172294","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T23:25:16.351949Z","strongest_claim":"Deployed at full traffic on Douyin, our system delivers significant improvements on key engagement metrics while meeting production latency, demonstrating a practical path to scaling end-to-end long-sequence recommendation to the 10k regime.","one_line_summary":"Douyin deploys stacked target-to-history cross attention and request-level batching to scale end-to-end recommendation modeling to 10k-length histories, observing scaling-law gains and live engagement improvements.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the length-extrapolative training strategy (train on shorter windows, infer on 10k) generalizes without performance loss and that stacked target-to-history cross attention captures the necessary preference signals from long histories.","pith_extraction_headline":"Stacked cross-attention and shared batching let recommendation models process 10,000-length user histories at production speed."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2511.06077/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":2,"snapshot_sha256":"7a41d85a461f01b63038805dbe2e9b71b94ca760b6a1a14637201f6949c8c6cd"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}