{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:IVHD2WKQ7FHMZACR7JNOGIWFDK","short_pith_number":"pith:IVHD2WKQ","schema_version":"1.0","canonical_sha256":"454e3d5950f94ecc8051fa5ae322c51a898c715e7ce145adebae4d5ae55608dc","source":{"kind":"arxiv","id":"2508.01174","version":1},"attestation_state":"computed","paper":{"title":"RSPO: Risk-Seeking Policy Optimization for Pass@k and Max@k Metrics in Large Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Haipeng Sun, Hong Dingqian, Hongfei Jiang, Hui Xiong, Junwei Bao, Kaichen Zhang, Shenghao Gao, Yang Song, Yuzhong Hong","submitted_at":"2025-08-02T03:25:26Z","abstract_excerpt":"Current large language model post-training optimizes a risk-neutral objective that maximizes expected reward, yet evaluation relies heavily on risk-seeking metrics like Pass@k (at least one success in k trials) and Max@k (maximum reward across k responses). This mismatch in risk preferences can inevitably lead to suboptimal performance. To bridge this gap, we propose Risk-Seeking Policy Optimization (RSPO), a novel method that directly targets Pass@k and Max@k during training. A key challenge in optimizing these metrics is the \"hitchhiking\" problem: low-reward responses are inadvertently reinf"},"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":"2508.01174","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-08-02T03:25:26Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"2fe19874f31f64bfb97ec7e8fb6e71782d998ad8f904ca2d5a7386610efbe238","abstract_canon_sha256":"a8f015ebe488313aac6ab21a0fca3f0d4eeeba2b824c257a28724cf54677afa5"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:47:34.244809Z","signature_b64":"qpvh66XpAoehAFPD71yyJ7oawDwJH12puMsPMNmxRBTAQi286frwCKfVAKY4ax2JAx5CDop6mLK31SzXVrtQCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"454e3d5950f94ecc8051fa5ae322c51a898c715e7ce145adebae4d5ae55608dc","last_reissued_at":"2026-07-05T11:47:34.244372Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:47:34.244372Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"RSPO: Risk-Seeking Policy Optimization for Pass@k and Max@k Metrics in Large Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Haipeng Sun, Hong Dingqian, Hongfei Jiang, Hui Xiong, Junwei Bao, Kaichen Zhang, Shenghao Gao, Yang Song, Yuzhong Hong","submitted_at":"2025-08-02T03:25:26Z","abstract_excerpt":"Current large language model post-training optimizes a risk-neutral objective that maximizes expected reward, yet evaluation relies heavily on risk-seeking metrics like Pass@k (at least one success in k trials) and Max@k (maximum reward across k responses). This mismatch in risk preferences can inevitably lead to suboptimal performance. To bridge this gap, we propose Risk-Seeking Policy Optimization (RSPO), a novel method that directly targets Pass@k and Max@k during training. A key challenge in optimizing these metrics is the \"hitchhiking\" problem: low-reward responses are inadvertently reinf"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2508.01174","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/2508.01174/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":"2508.01174","created_at":"2026-07-05T11:47:34.244430+00:00"},{"alias_kind":"arxiv_version","alias_value":"2508.01174v1","created_at":"2026-07-05T11:47:34.244430+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2508.01174","created_at":"2026-07-05T11:47:34.244430+00:00"},{"alias_kind":"pith_short_12","alias_value":"IVHD2WKQ7FHM","created_at":"2026-07-05T11:47:34.244430+00:00"},{"alias_kind":"pith_short_16","alias_value":"IVHD2WKQ7FHMZACR","created_at":"2026-07-05T11:47:34.244430+00:00"},{"alias_kind":"pith_short_8","alias_value":"IVHD2WKQ","created_at":"2026-07-05T11:47:34.244430+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.17333","citing_title":"Leveraging Error Diversity in Group Rollouts for Reinforcement Learning","ref_index":33,"is_internal_anchor":false},{"citing_arxiv_id":"2605.17333","citing_title":"Leveraging Error Diversity in Group Rollouts for Reinforcement Learning","ref_index":33,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/IVHD2WKQ7FHMZACR7JNOGIWFDK","json":"https://pith.science/pith/IVHD2WKQ7FHMZACR7JNOGIWFDK.json","graph_json":"https://pith.science/api/pith-number/IVHD2WKQ7FHMZACR7JNOGIWFDK/graph.json","events_json":"https://pith.science/api/pith-number/IVHD2WKQ7FHMZACR7JNOGIWFDK/events.json","paper":"https://pith.science/paper/IVHD2WKQ"},"agent_actions":{"view_html":"https://pith.science/pith/IVHD2WKQ7FHMZACR7JNOGIWFDK","download_json":"https://pith.science/pith/IVHD2WKQ7FHMZACR7JNOGIWFDK.json","view_paper":"https://pith.science/paper/IVHD2WKQ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2508.01174&json=true","fetch_graph":"https://pith.science/api/pith-number/IVHD2WKQ7FHMZACR7JNOGIWFDK/graph.json","fetch_events":"https://pith.science/api/pith-number/IVHD2WKQ7FHMZACR7JNOGIWFDK/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/IVHD2WKQ7FHMZACR7JNOGIWFDK/action/timestamp_anchor","attest_storage":"https://pith.science/pith/IVHD2WKQ7FHMZACR7JNOGIWFDK/action/storage_attestation","attest_author":"https://pith.science/pith/IVHD2WKQ7FHMZACR7JNOGIWFDK/action/author_attestation","sign_citation":"https://pith.science/pith/IVHD2WKQ7FHMZACR7JNOGIWFDK/action/citation_signature","submit_replication":"https://pith.science/pith/IVHD2WKQ7FHMZACR7JNOGIWFDK/action/replication_record"}},"created_at":"2026-07-05T11:47:34.244430+00:00","updated_at":"2026-07-05T11:47:34.244430+00:00"}