{ "profile": "wrs-published-research-view-1", "edition": "ratings-e039106f30dec177", "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS", "entry": { "id": "fu-et-al-areal-asynchronous-rl-2025", "title": "AReaL: asynchronous rollouts with staleness-aware policy updates", "summary": "Decoupled continuous generation from training, using interruptible rollouts, bounded data staleness, dynamic batching and a decoupled PPO objective for trajectories generated across policy versions.", "boundary": "Inherits asynchronous RL and decoupled PPO. Weight updates can interrupt generation, discard/recompute KV caches and create mixed-policy trajectories; staleness must be constrained, not ignored. Up to2.57fold acceleration and matched accuracy concern selected math/code experiments, not universal convergence or arbitrary off-policy stability.", "area": "Distributed language-model reinforcement learning", "topic": "Artificial intelligence", "year": 2025, "lane": "HUMAN", "result": "ALGORITHM", "package_id": "fu-et-al-areal-asynchronous-rl-2025", "target": { "record_id": "urn:uuid:3ee023a4-297a-42ee-a2ad-4e6a378ade70", "version": 1, "content_hash": "sha256:b426e9e6a1a25501d8d64caeb8d18682992f5e74496d7f9c930b89687eefc090" }, "humans": [ "Wei Fu", "Jiaxuan Gao", "Xujie Shen", "Chen Zhu", "Zhiyu Mei", "Chuyi He", "Shusheng Xu", "Guo Wei", "Jun Mei", "Jiashu Wang", "Tongkai Yang", "Binhang Yuan", "Yi Wu" ], "authors_complete": true, "human_role_note": "Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.", "agents": [], "related": [], "subresults": [], "scientific_assessment": { "source_review": "SOURCE_CURATED", "scientific_review": "NOT_REVIEWED_BY_WRS", "formal_replay": "NOT_RUN", "provider_attestation": "NONE" }, "sources": [ { "id": "cur20261015-arxiv-2505-24298v1", "title": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1" } ], "points": "75000", "published_grade": { "id": "fu-et-al-areal-asynchronous-rl-2025", "entry_id": "fu-et-al-areal-asynchronous-rl-2025", "band": "B", "modifier": "+", "target": { "record_id": "urn:uuid:3ee023a4-297a-42ee-a2ad-4e6a378ade70", "version": 1, "content_hash": "sha256:b426e9e6a1a25501d8d64caeb8d18682992f5e74496d7f9c930b89687eefc090" }, "what_changed": "Decoupled continuous generation from training, using interruptible rollouts, bounded data staleness, dynamic batching and a decoupled PPO objective for trajectories generated across policy versions.", "scope": "Decoupled continuous generation from training, using interruptible rollouts, bounded data staleness, dynamic batching and a decoupled PPO objective for trajectories generated across policy versions. Inherits asynchronous RL and decoupled PPO. Weight updates can interrupt generation, discard/recompute KV caches and create mixed-policy trajectories; staleness must be constrained, not ignored. Up to2.57fold acceleration and matched accuracy concern selected math/code experiments, not universal convergence or arbitrary off-policy stability.", "status": "PROVISIONAL_EDITORIAL", "scientific_correctness_certified": false }, "assessment_disposition": { "status": "ASSESSED", "label": "Provisionally assessed", "reason": "A current grade and separate reference score identify this exact contribution.", "target": { "record_id": "urn:uuid:3ee023a4-297a-42ee-a2ad-4e6a378ade70", "version": 1, "content_hash": "sha256:b426e9e6a1a25501d8d64caeb8d18682992f5e74496d7f9c930b89687eefc090" }, "edition": "ratings-e039106f30dec177" } }, "version": { "kind": "PUBLIC_RECORD_VIEW", "previous": null, "published_at": "2026-10-16T00:00:00Z", "publisher": { "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33", "role": "WRS_EDITORIAL_PUBLICATION" }, "record": { "contributors": [ { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Wei Fu", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Jiaxuan Gao", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Xujie Shen", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Chen Zhu", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Zhiyu Mei", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Chuyi He", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Shusheng Xu", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Guo Wei", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Jun Mei", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Jiashu Wang", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Tongkai Yang", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Binhang Yuan", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Yi Wu", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1" } } ], "evidence": [ { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-0", "kind": "source_reference", "source": { "citation": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1" } } ], "origin": "HUMAN_LEGACY", "record_profile": "wrs-local-research-notes-0.1", "record_type": "research_note", "rrp_version": "0.1", "scope": { "description": "Inherits asynchronous RL and decoupled PPO. Weight updates can interrupt generation, discard/recompute KV caches and create mixed-policy trajectories; staleness must be constrained, not ignored. Up to2.57fold acceleration and matched accuracy concern selected math/code experiments, not universal convergence or arbitrary off-policy stability.", "domain": "Distributed language-model reinforcement learning" }, "sources": [ { "citation": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1" } ], "statement": "# AReaL: asynchronous rollouts with staleness-aware policy updates\n\nDecoupled continuous generation from training, using interruptible rollouts, bounded data staleness, dynamic batching and a decoupled PPO objective for trajectories generated across policy versions.\n\n## Scientific scope\n\nInherits asynchronous RL and decoupled PPO. Weight updates can interrupt generation, discard/recompute KV caches and create mixed-policy trajectories; staleness must be constrained, not ignored. Up to2.57fold acceleration and matched accuracy concern selected math/code experiments, not universal convergence or arbitrary off-policy stability.\n\n## Reading status\n\nThis is a source-linked WRS catalogue note. A full explanatory guide and independent scientific reconstruction have not been completed for this entry. Read the identified original work for its complete narrative and arguments.\n\n## People and systems\n\nWei Fu; Jiaxuan Gao; Xujie Shen; Chen Zhu; Zhiyu Mei; Chuyi He; Shusheng Xu; Guo Wei; Jun Mei; Jiashu Wang; Tongkai Yang; Binhang Yuan; Yi Wu\n\nSource-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.\n\n\n\n## Original sources\n\n- [AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning](https://arxiv.org/abs/2505.24298v1)\n", "summary": "Decoupled continuous generation from training, using interruptible rollouts, bounded data staleness, dynamic batching and a decoupled PPO objective for trajectories generated across policy versions.", "title": "AReaL: asynchronous rollouts with staleness-aware policy updates" }, "record_id": "urn:uuid:3ee023a4-297a-42ee-a2ad-4e6a378ade70", "version": 1, "source_ref": { "record_id": "urn:uuid:3ee023a4-297a-42ee-a2ad-4e6a378ade70", "version": 1, "content_hash": "sha256:b426e9e6a1a25501d8d64caeb8d18682992f5e74496d7f9c930b89687eefc090" }, "export_view": "wrs-public-editorial-view-1", "public_view_hash": "sha256:a56d44991e6700b09f45e7cabed3898c92a769127634d7b6ab6b3fb58228c546" }, "target": { "record_id": "urn:uuid:3ee023a4-297a-42ee-a2ad-4e6a378ade70", "version": 1, "content_hash": "sha256:b426e9e6a1a25501d8d64caeb8d18682992f5e74496d7f9c930b89687eefc090" }, "published_grade": { "profile": "wrs-published-grade-1", "edition": "2026-10-28.1", "id": "fu-et-al-areal-asynchronous-rl-2025", "entry_id": "fu-et-al-areal-asynchronous-rl-2025", "revision": 1, "title": "AReaL: asynchronous rollouts with staleness-aware policy updates", "context": "Distributed language-model reinforcement learning", "track": "Human-designed research method or resource", "band": "B", "modifier": "+", "status": "PROVISIONAL_EDITORIAL", "assessed_on": "2026-10-16", "evidence_cutoff": "2026-10-16", "achievement_state": "SUPPORTED_SCOPED_RESULT", "what_changed": "Decoupled continuous generation from training, using interruptible rollouts, bounded data staleness, dynamic batching and a decoupled PPO objective for trajectories generated across policy versions.", "scope": "Decoupled continuous generation from training, using interruptible rollouts, bounded data staleness, dynamic batching and a decoupled PPO objective for trajectories generated across policy versions. Inherits asynchronous RL and decoupled PPO. Weight updates can interrupt generation, discard/recompute KV caches and create mixed-policy trajectories; staleness must be constrained, not ignored. Up to2.57fold acceleration and matched accuracy concern selected math/code experiments, not universal convergence or arbitrary off-policy stability.", "evidence_summary": "The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Inherits asynchronous RL and decoupled PPO. Weight updates can interrupt generation, discard/recompute KV caches and create mixed-policy trajectories; staleness must be constrained, not ignored. Up to2.57fold acceleration and matched accuracy concern selected math/code experiments, not universal convergence or arbitrary off-policy stability.", "references": [ { "title": "AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning", "url": "https://arxiv.org/abs/2505.24298v1", "locator": "Original contribution and stated assumptions, evaluation scope and limitations; source-linked assessment, not independent execution or proof certification." } ], "roles": [ { "who": "Source-listed authors and technical contributors", "credit": "SCIENCE", "did": "Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities." } ], "target": { "record_id": "urn:uuid:3ee023a4-297a-42ee-a2ad-4e6a378ade70", "version": 1, "content_hash": "sha256:b426e9e6a1a25501d8d64caeb8d18682992f5e74496d7f9c930b89687eefc090" }, "scientific_correctness_certified": false, "independent_expert_calibration": false, "method": "PROPRIETARY", "points": "75000" }, "canonical_agent_submission": false, "scientific_verification_awarded": false, "reader_guide": null, "reading_status": "SOURCE_LINKED_NOTE" }