{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:V2OKT5MH3OVVKOFUGY62YTYXLF","short_pith_number":"pith:V2OKT5MH","schema_version":"1.0","canonical_sha256":"ae9ca9f587dbab5538b4363dac4f1759636b5ec3419d8f5f5f7ff065002b6f9f","source":{"kind":"arxiv","id":"2606.00395","version":1},"attestation_state":"computed","paper":{"title":"PR2: Predictive Routing Replay for MoE-Based LLM Reinforcement Learning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Daize Dong, Dimitris N. Metaxas, Emad Barsoum, Haolong Jia, Hongyi Wang, Huanwei Di, Jialian Wu, Jiang Liu, Jiawei Wu, Junlin Chen, Zhengzhong Liu, Zicheng Liu","submitted_at":"2026-05-29T22:28:08Z","abstract_excerpt":"Mixture of Experts (MoE) Large Language Models (LLMs) achieve strong performance at scale. However, reinforcement learning (RL) on MoE-based LLMs often suffers from training instability. A root cause is router drift, i.e., expert activations can change drastically across model updates and differ between disaggregated rollout and training phases, causing large rollout--training mismatch and unstable importance sampling weights in PPO-style RL algorithms. Routing replay mitigates this issue by freezing the replay route within each reasoning trajectory, but it ignores how the router evolves under"},"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":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2606.00395","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-05-29T22:28:08Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"123a0757d1e2396ebe14bfdfff822513bbf8b70a7a36fe25e0e6a987a45c9bd6","abstract_canon_sha256":"046b8db29b22d419488d515b855cca22a2d640b12b4316f87cdd13f3a50dc74a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-02T01:03:53.547645Z","signature_b64":"0h40Jq6RhNRndDs6Yji1q7375G9oGCBSYm/5ze2C9jLWkwDCb6HotVCL1zEwy9AvQTRA/qmAJlQz8ymYjCdwBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"ae9ca9f587dbab5538b4363dac4f1759636b5ec3419d8f5f5f7ff065002b6f9f","last_reissued_at":"2026-06-02T01:03:53.547216Z","signature_status":"signed_v1","first_computed_at":"2026-06-02T01:03:53.547216Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"PR2: Predictive Routing Replay for MoE-Based LLM Reinforcement Learning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Daize Dong, Dimitris N. Metaxas, Emad Barsoum, Haolong Jia, Hongyi Wang, Huanwei Di, Jialian Wu, Jiang Liu, Jiawei Wu, Junlin Chen, Zhengzhong Liu, Zicheng Liu","submitted_at":"2026-05-29T22:28:08Z","abstract_excerpt":"Mixture of Experts (MoE) Large Language Models (LLMs) achieve strong performance at scale. However, reinforcement learning (RL) on MoE-based LLMs often suffers from training instability. A root cause is router drift, i.e., expert activations can change drastically across model updates and differ between disaggregated rollout and training phases, causing large rollout--training mismatch and unstable importance sampling weights in PPO-style RL algorithms. Routing replay mitigates this issue by freezing the replay route within each reasoning trajectory, but it ignores how the router evolves under"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2606.00395","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2606.00395/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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2606.00395","created_at":"2026-06-02T01:03:53.547297+00:00"},{"alias_kind":"arxiv_version","alias_value":"2606.00395v1","created_at":"2026-06-02T01:03:53.547297+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2606.00395","created_at":"2026-06-02T01:03:53.547297+00:00"},{"alias_kind":"pith_short_12","alias_value":"V2OKT5MH3OVV","created_at":"2026-06-02T01:03:53.547297+00:00"},{"alias_kind":"pith_short_16","alias_value":"V2OKT5MH3OVVKOFU","created_at":"2026-06-02T01:03:53.547297+00:00"},{"alias_kind":"pith_short_8","alias_value":"V2OKT5MH","created_at":"2026-06-02T01:03:53.547297+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/V2OKT5MH3OVVKOFUGY62YTYXLF","json":"https://pith.science/pith/V2OKT5MH3OVVKOFUGY62YTYXLF.json","graph_json":"https://pith.science/api/pith-number/V2OKT5MH3OVVKOFUGY62YTYXLF/graph.json","events_json":"https://pith.science/api/pith-number/V2OKT5MH3OVVKOFUGY62YTYXLF/events.json","paper":"https://pith.science/paper/V2OKT5MH"},"agent_actions":{"view_html":"https://pith.science/pith/V2OKT5MH3OVVKOFUGY62YTYXLF","download_json":"https://pith.science/pith/V2OKT5MH3OVVKOFUGY62YTYXLF.json","view_paper":"https://pith.science/paper/V2OKT5MH","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2606.00395&json=true","fetch_graph":"https://pith.science/api/pith-number/V2OKT5MH3OVVKOFUGY62YTYXLF/graph.json","fetch_events":"https://pith.science/api/pith-number/V2OKT5MH3OVVKOFUGY62YTYXLF/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/V2OKT5MH3OVVKOFUGY62YTYXLF/action/timestamp_anchor","attest_storage":"https://pith.science/pith/V2OKT5MH3OVVKOFUGY62YTYXLF/action/storage_attestation","attest_author":"https://pith.science/pith/V2OKT5MH3OVVKOFUGY62YTYXLF/action/author_attestation","sign_citation":"https://pith.science/pith/V2OKT5MH3OVVKOFUGY62YTYXLF/action/citation_signature","submit_replication":"https://pith.science/pith/V2OKT5MH3OVVKOFUGY62YTYXLF/action/replication_record"}},"created_at":"2026-06-02T01:03:53.547297+00:00","updated_at":"2026-06-02T01:03:53.547297+00:00"}