{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:JDURUAVDSXWPBLCSDSGB7KZR2R","short_pith_number":"pith:JDURUAVD","schema_version":"1.0","canonical_sha256":"48e91a02a395ecf0ac521c8c1fab31d451214957c417e859b466b8474f3f7860","source":{"kind":"arxiv","id":"2605.14278","version":1},"attestation_state":"computed","paper":{"title":"KVPO: ODE-Native GRPO for Autoregressive Video Alignment via KV Semantic Exploration","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"KVPO aligns streaming autoregressive video generators by routing semantic variations through the KV cache and modeling policies via velocity energy in ODE space.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Jun Zhou, Kaixi Cong, Ruicheng Zhang, Shuiyang Mao, Wei Liu, Xiu Li, Zhizhou Zhong, Zunnan Xu","submitted_at":"2026-05-14T02:24:46Z","abstract_excerpt":"Aligning streaming autoregressive (AR) video generators with human preferences is challenging. Existing reinforcement learning methods predominantly rely on noise-based exploration and SDE-based surrogate policies that are mismatched to the deterministic ODE dynamics of distilled AR models, and tend to perturb low-level appearance rather than the high-level semantic storyline progression critical for long-horizon coherence. To address these limitations, we present KVPO, an ODE-native online Group Relative Policy Optimization (GRPO) framework for aligning streaming video generators. For diversi"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":true,"formal_links_present":false},"canonical_record":{"source":{"id":"2605.14278","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-05-14T02:24:46Z","cross_cats_sorted":[],"title_canon_sha256":"b2a8e643e00a2a42fde144366c302edf531898314362e5fc7e91e727fb8a3b93","abstract_canon_sha256":"593bffff0e2365eaac80c1b925b27ac35f20ae5568ddceb7175a0ec86eda6eb5"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:39:10.316796Z","signature_b64":"vMzS3iQFvsTarYte0YWLROtuM1C2uUoI2WDmHqU4+V8n3C7N5JBRxCQZbilVK192kkSBenvQmwYf4JQD04XJCQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"48e91a02a395ecf0ac521c8c1fab31d451214957c417e859b466b8474f3f7860","last_reissued_at":"2026-05-17T23:39:10.316265Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:39:10.316265Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"KVPO: ODE-Native GRPO for Autoregressive Video Alignment via KV Semantic Exploration","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"KVPO aligns streaming autoregressive video generators by routing semantic variations through the KV cache and modeling policies via velocity energy in ODE space.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Jun Zhou, Kaixi Cong, Ruicheng Zhang, Shuiyang Mao, Wei Liu, Xiu Li, Zhizhou Zhong, Zunnan Xu","submitted_at":"2026-05-14T02:24:46Z","abstract_excerpt":"Aligning streaming autoregressive (AR) video generators with human preferences is challenging. Existing reinforcement learning methods predominantly rely on noise-based exploration and SDE-based surrogate policies that are mismatched to the deterministic ODE dynamics of distilled AR models, and tend to perturb low-level appearance rather than the high-level semantic storyline progression critical for long-horizon coherence. To address these limitations, we present KVPO, an ODE-native online Group Relative Policy Optimization (GRPO) framework for aligning streaming video generators. For diversi"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"KVPO introduces a causal-semantic exploration paradigm that relocates the source of variation from stochastic noise to the historical KV cache, constructing semantically diverse generation branches that remain strictly on the data manifold, and a velocity-field surrogate policy based on Trajectory Velocity Energy that yields a reward-weighted contrastive objective fully consistent with the native ODE formulation.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That stochastically routing historical KV entries produces semantically diverse generation branches that remain strictly on the data manifold without introducing artifacts or deviating from the model's learned distribution.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"KVPO aligns streaming autoregressive video generators with human preferences via ODE-native GRPO, using KV cache for semantic exploration and TVE for velocity-based policy modeling, yielding gains in quality and alignment.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"KVPO aligns streaming autoregressive video generators by routing semantic variations through the KV cache and modeling policies via velocity energy in ODE space.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"3854c0bab6a8440eef79302105917c1f916ce464dae5b55e440e45335c16d341"},"source":{"id":"2605.14278","kind":"arxiv","version":1},"verdict":{"id":"f7558614-696f-4ed3-87a9-cce9d31f28a9","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T02:35:53.482864Z","strongest_claim":"KVPO introduces a causal-semantic exploration paradigm that relocates the source of variation from stochastic noise to the historical KV cache, constructing semantically diverse generation branches that remain strictly on the data manifold, and a velocity-field surrogate policy based on Trajectory Velocity Energy that yields a reward-weighted contrastive objective fully consistent with the native ODE formulation.","one_line_summary":"KVPO aligns streaming autoregressive video generators with human preferences via ODE-native GRPO, using KV cache for semantic exploration and TVE for velocity-based policy modeling, yielding gains in quality and alignment.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That stochastically routing historical KV entries produces semantically diverse generation branches that remain strictly on the data manifold without introducing artifacts or deviating from the model's learned distribution.","pith_extraction_headline":"KVPO aligns streaming autoregressive video generators by routing semantic variations through the KV cache and modeling policies via velocity energy in ODE space."},"references":{"count":33,"sample":[{"doi":"","year":2025,"title":"arXiv preprint arXiv:2511.16955 (2025) 3","work_id":"20b57bf7-3f3a-4078-a74b-d8065c609465","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"arXiv preprint arXiv:2603.17461 (2026)","work_id":"20db7b14-851b-4a3b-9b8a-53fe5fca0290","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"arXiv preprint arXiv:2603.21299 (2026)","work_id":"d62c2c5a-3956-4522-bb31-438b3c66c6c5","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"LoRA: Low-Rank Adaptation of Large Language Models","work_id":"0426219a-789e-4964-adc8-a04538510818","ref_index":4,"cited_arxiv_id":"2106.09685","is_internal_anchor":true},{"doi":"","year":2025,"title":"Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion","work_id":"53e58ef9-7932-4b83-b757-34ac14db3e0f","ref_index":5,"cited_arxiv_id":"2506.08009","is_internal_anchor":true}],"resolved_work":33,"snapshot_sha256":"536eea658fb28bb4fd023b78b136fe998ec051f75f2dc0fa7045ea4cff6b43d6","internal_anchors":12},"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2605.14278","created_at":"2026-05-17T23:39:10.316363+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.14278v1","created_at":"2026-05-17T23:39:10.316363+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.14278","created_at":"2026-05-17T23:39:10.316363+00:00"},{"alias_kind":"pith_short_12","alias_value":"JDURUAVDSXWP","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"JDURUAVDSXWPBLCS","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"JDURUAVD","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2605.15803","citing_title":"Embedding-perturbed Exploration Preference Optimization for Flow Models","ref_index":94,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/JDURUAVDSXWPBLCSDSGB7KZR2R","json":"https://pith.science/pith/JDURUAVDSXWPBLCSDSGB7KZR2R.json","graph_json":"https://pith.science/api/pith-number/JDURUAVDSXWPBLCSDSGB7KZR2R/graph.json","events_json":"https://pith.science/api/pith-number/JDURUAVDSXWPBLCSDSGB7KZR2R/events.json","paper":"https://pith.science/paper/JDURUAVD"},"agent_actions":{"view_html":"https://pith.science/pith/JDURUAVDSXWPBLCSDSGB7KZR2R","download_json":"https://pith.science/pith/JDURUAVDSXWPBLCSDSGB7KZR2R.json","view_paper":"https://pith.science/paper/JDURUAVD","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.14278&json=true","fetch_graph":"https://pith.science/api/pith-number/JDURUAVDSXWPBLCSDSGB7KZR2R/graph.json","fetch_events":"https://pith.science/api/pith-number/JDURUAVDSXWPBLCSDSGB7KZR2R/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/JDURUAVDSXWPBLCSDSGB7KZR2R/action/timestamp_anchor","attest_storage":"https://pith.science/pith/JDURUAVDSXWPBLCSDSGB7KZR2R/action/storage_attestation","attest_author":"https://pith.science/pith/JDURUAVDSXWPBLCSDSGB7KZR2R/action/author_attestation","sign_citation":"https://pith.science/pith/JDURUAVDSXWPBLCSDSGB7KZR2R/action/citation_signature","submit_replication":"https://pith.science/pith/JDURUAVDSXWPBLCSDSGB7KZR2R/action/replication_record"}},"created_at":"2026-05-17T23:39:10.316363+00:00","updated_at":"2026-05-17T23:39:10.316363+00:00"}