{"paper":{"title":"HorizonDrive: Self-Corrective Autoregressive World Model for Long-horizon Driving Simulation","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"A self-corrective training procedure allows autoregressive driving models to generate minute-scale simulations without drift.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Conglang Zhang, Qian Zhang, Qingjie Wang, Weiqiang Ren, Wei Yin, Xiaoyang Guo, Yifan Zhan, Yinqiang Zheng, Yu Li, Zhanpeng Ouyang, Zhen Dong, Zhengqing Chen, Zihao Yang","submitted_at":"2026-05-12T06:22:16Z","abstract_excerpt":"Closed-loop driving simulation requires real-time interaction beyond short offline clips, pushing current driving world models toward autoregressive (AR) rollout. Existing AR distillation approaches typically rely on frame sinks or student-side degradation training. The former transfers poorly to driving due to fast ego-motion and rapid scene changes, while the latter remains bounded by the teacher's single-pass output length and thus provides only a limited supervision horizon. A natural question is: can the teacher itself be extended via AR rollout to provide unbounded-horizon supervision at"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"HorizonDrive natively supports minute-scale AR rollout under bounded memory; on nuScenes, HorizonDrive reduces FID by 52% and FVD by 37%, and lowers ARE and DTW by 21% and 9% relative to the strongest long-horizon streaming baselines, while remaining competitive with single-pass driving video generators.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That training with scheduled rollout recovery produces a teacher model that remains stable and provides reliable supervision across long autoregressive rollouts without introducing new biases or artifacts not present in ground truth.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"HorizonDrive enables stable long-horizon autoregressive driving simulation via anti-drifting teacher training with scheduled rollout recovery and teacher rollout distillation.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A self-corrective training procedure allows autoregressive driving models to generate minute-scale simulations without drift.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"2a80e299f8a0859887f0b1630f4237f65d4f263b12f3635ca9bec815622c0d5e"},"source":{"id":"2605.11596","kind":"arxiv","version":2},"verdict":{"id":"316ef20e-919e-40be-9c5d-b6dc7fe84031","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-13T01:12:39.233806Z","strongest_claim":"HorizonDrive natively supports minute-scale AR rollout under bounded memory; on nuScenes, HorizonDrive reduces FID by 52% and FVD by 37%, and lowers ARE and DTW by 21% and 9% relative to the strongest long-horizon streaming baselines, while remaining competitive with single-pass driving video generators.","one_line_summary":"HorizonDrive enables stable long-horizon autoregressive driving simulation via anti-drifting teacher training with scheduled rollout recovery and teacher rollout distillation.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That training with scheduled rollout recovery produces a teacher model that remains stable and provides reliable supervision across long autoregressive rollouts without introducing new biases or artifacts not present in ground truth.","pith_extraction_headline":"A self-corrective training procedure allows autoregressive driving models to generate minute-scale simulations without drift."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.11596/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-21T01:01:33.499523Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-20T15:54:54.973651Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-20T04:02:00.490529Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T12:33:25.843749Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"0dd64dfb3edc119fce7af09a02d03e128fcfd73489dc7c257b49e366115a8d83"},"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"}