{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:5T3JCLYXK3UK6MSXLGSJEUWPXL","short_pith_number":"pith:5T3JCLYX","schema_version":"1.0","canonical_sha256":"ecf6912f1756e8af325759a49252cfbaef349a6d5a9e0e4e2b22fa8b61802a22","source":{"kind":"arxiv","id":"2605.12058","version":2},"attestation_state":"computed","paper":{"title":"Holder Policy Optimisation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"HölderPO resolves GRPO's aggregation trade-off by using a tunable Hölder mean with annealed parameter p to control gradient concentration and variance.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Chenyang Le, Dingli Liang, Jiachen Zhu, Jianghao Lin, Jun Wang, Lingyu Yang, Weinan Zhang, Yihang Chen, Yuxiang Chen, Zhaokai Wang, Ziqin Gong","submitted_at":"2026-05-12T12:45:03Z","abstract_excerpt":"Group Relative Policy Optimisation (GRPO) enhances large language models by estimating advantages across a group of sampled trajectories. However, mapping these trajectory-level advantages to policy updates requires aggregating token-level probabilities within each sequence. Relying on a fixed aggregation mechanism for this step fundamentally limits the algorithm's adaptability. Empirically, we observe a critical trade-off: certain fixed aggregations frequently suffer from training collapse, while others fail to yield satisfactory performance. To resolve this, we propose \\textbf{H\\\"{o}lderPO},"},"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":"2605.12058","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-12T12:45:03Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"69378e129fc2f4f34aa8ffab9c29b8251f76273548881c5cac738d2b94b74a9a","abstract_canon_sha256":"f74a957ad172eea0f389658e2d4e70c908f04e77b1b0ae047729822bbb14d558"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-22T01:04:06.183707Z","signature_b64":"eG1QT2ujXloQFnGpqYq2myxlZtXcfqBpsQ+qdQQSzZHAQNPw3Donp7V2noNKc8urlwNWLjkIKBZFoEtMZ4T4Cg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"ecf6912f1756e8af325759a49252cfbaef349a6d5a9e0e4e2b22fa8b61802a22","last_reissued_at":"2026-05-22T01:04:06.183005Z","signature_status":"signed_v1","first_computed_at":"2026-05-22T01:04:06.183005Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Holder Policy Optimisation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"HölderPO resolves GRPO's aggregation trade-off by using a tunable Hölder mean with annealed parameter p to control gradient concentration and variance.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Chenyang Le, Dingli Liang, Jiachen Zhu, Jianghao Lin, Jun Wang, Lingyu Yang, Weinan Zhang, Yihang Chen, Yuxiang Chen, Zhaokai Wang, Ziqin Gong","submitted_at":"2026-05-12T12:45:03Z","abstract_excerpt":"Group Relative Policy Optimisation (GRPO) enhances large language models by estimating advantages across a group of sampled trajectories. However, mapping these trajectory-level advantages to policy updates requires aggregating token-level probabilities within each sequence. Relying on a fixed aggregation mechanism for this step fundamentally limits the algorithm's adaptability. Empirically, we observe a critical trade-off: certain fixed aggregations frequently suffer from training collapse, while others fail to yield satisfactory performance. To resolve this, we propose \\textbf{H\\\"{o}lderPO},"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"our approach achieves a state-of-the-art average accuracy of 54.9% across multiple mathematical benchmarks, yielding a substantial 7.2% relative gain over standard GRPO and secures an exceptional 93.8% success rate on ALFWorld.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That modulating the single scalar p via annealing will reliably resolve the concentration-stability trade-off across different model sizes, tasks, and sampling budgets without introducing new failure modes not captured by the reported experiments.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"HölderPO unifies token aggregation in GRPO via the Hölder mean with dynamic p annealing, reporting 54.9% average math-benchmark accuracy and 93.8% ALFWorld success.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"HölderPO resolves GRPO's aggregation trade-off by using a tunable Hölder mean with annealed parameter p to control gradient concentration and variance.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"921a14289d15c434d5c9bf698699e05d167af22f773ed5a22ff67dea8289679d"},"source":{"id":"2605.12058","kind":"arxiv","version":2},"verdict":{"id":"1c02ce04-4baa-4742-adac-638daa1a1dd5","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-13T05:58:47.841499Z","strongest_claim":"our approach achieves a state-of-the-art average accuracy of 54.9% across multiple mathematical benchmarks, yielding a substantial 7.2% relative gain over standard GRPO and secures an exceptional 93.8% success rate on ALFWorld.","one_line_summary":"HölderPO unifies token aggregation in GRPO via the Hölder mean with dynamic p annealing, reporting 54.9% average math-benchmark accuracy and 93.8% ALFWorld success.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That modulating the single scalar p via annealing will reliably resolve the concentration-stability trade-off across different model sizes, tasks, and sampling budgets without introducing new failure modes not captured by the reported experiments.","pith_extraction_headline":"HölderPO resolves GRPO's aggregation trade-off by using a tunable Hölder mean with annealed parameter p to control gradient concentration and variance."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.12058/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-20T17:01:26.385247Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-20T11:25:11.599221Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T23:01:58.411205Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T11:33:48.696831Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"5f5ffa52975d0f31002e74ed111dc23a82e862fd4809b4cbbae06f482fcc3429"},"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":"2605.12058","created_at":"2026-05-22T01:04:06.183096+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.12058v2","created_at":"2026-05-22T01:04:06.183096+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.12058","created_at":"2026-05-22T01:04:06.183096+00:00"},{"alias_kind":"pith_short_12","alias_value":"5T3JCLYXK3UK","created_at":"2026-05-22T01:04:06.183096+00:00"},{"alias_kind":"pith_short_16","alias_value":"5T3JCLYXK3UK6MSX","created_at":"2026-05-22T01:04:06.183096+00:00"},{"alias_kind":"pith_short_8","alias_value":"5T3JCLYX","created_at":"2026-05-22T01:04:06.183096+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/5T3JCLYXK3UK6MSXLGSJEUWPXL","json":"https://pith.science/pith/5T3JCLYXK3UK6MSXLGSJEUWPXL.json","graph_json":"https://pith.science/api/pith-number/5T3JCLYXK3UK6MSXLGSJEUWPXL/graph.json","events_json":"https://pith.science/api/pith-number/5T3JCLYXK3UK6MSXLGSJEUWPXL/events.json","paper":"https://pith.science/paper/5T3JCLYX"},"agent_actions":{"view_html":"https://pith.science/pith/5T3JCLYXK3UK6MSXLGSJEUWPXL","download_json":"https://pith.science/pith/5T3JCLYXK3UK6MSXLGSJEUWPXL.json","view_paper":"https://pith.science/paper/5T3JCLYX","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.12058&json=true","fetch_graph":"https://pith.science/api/pith-number/5T3JCLYXK3UK6MSXLGSJEUWPXL/graph.json","fetch_events":"https://pith.science/api/pith-number/5T3JCLYXK3UK6MSXLGSJEUWPXL/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/5T3JCLYXK3UK6MSXLGSJEUWPXL/action/timestamp_anchor","attest_storage":"https://pith.science/pith/5T3JCLYXK3UK6MSXLGSJEUWPXL/action/storage_attestation","attest_author":"https://pith.science/pith/5T3JCLYXK3UK6MSXLGSJEUWPXL/action/author_attestation","sign_citation":"https://pith.science/pith/5T3JCLYXK3UK6MSXLGSJEUWPXL/action/citation_signature","submit_replication":"https://pith.science/pith/5T3JCLYXK3UK6MSXLGSJEUWPXL/action/replication_record"}},"created_at":"2026-05-22T01:04:06.183096+00:00","updated_at":"2026-05-22T01:04:06.183096+00:00"}