{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:J47MNN3MCJ56E36HF4TAA7YY46","short_pith_number":"pith:J47MNN3M","schema_version":"1.0","canonical_sha256":"4f3ec6b76c127be26fc72f26007f18e79155ce1c16c9137b7945786637a225ad","source":{"kind":"arxiv","id":"2503.01333","version":1},"attestation_state":"computed","paper":{"title":"Group Relative Policy Optimization for Image Captioning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Xu Liang","submitted_at":"2025-03-03T09:16:41Z","abstract_excerpt":"Image captioning tasks usually use two-stage training to complete model optimization. The first stage uses cross-entropy as the loss function for optimization, and the second stage uses self-critical sequence training (SCST) for reinforcement learning optimization. However, the SCST algorithm has certain defects. SCST relies only on a single greedy decoding result as a baseline. If the model itself is not stable enough, the greedy decoding result may be relatively worst, which will lead to a high variance of advantage estimation, further leading to unstable policy updates. In addition, SCST on"},"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":"2503.01333","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2025-03-03T09:16:41Z","cross_cats_sorted":[],"title_canon_sha256":"51d81ca7185d1e4724287358ec801cde80914d4110f2e7ef54da16faea802a4b","abstract_canon_sha256":"9dda44f8f7410db58518982c2ccde31b9bdcc5564b8e031147496650d35eaa43"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:23:04.798854Z","signature_b64":"tBbwxFqkMieea/fh+//av037lfixGCfRwYSsrYGDJmS8lCcMhNCjzCBQ3M/5/QWGbjEqSoCB3QMPsZicaWXVDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4f3ec6b76c127be26fc72f26007f18e79155ce1c16c9137b7945786637a225ad","last_reissued_at":"2026-07-05T10:23:04.798063Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:23:04.798063Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Group Relative Policy Optimization for Image Captioning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Xu Liang","submitted_at":"2025-03-03T09:16:41Z","abstract_excerpt":"Image captioning tasks usually use two-stage training to complete model optimization. The first stage uses cross-entropy as the loss function for optimization, and the second stage uses self-critical sequence training (SCST) for reinforcement learning optimization. However, the SCST algorithm has certain defects. SCST relies only on a single greedy decoding result as a baseline. If the model itself is not stable enough, the greedy decoding result may be relatively worst, which will lead to a high variance of advantage estimation, further leading to unstable policy updates. In addition, SCST on"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2503.01333","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/2503.01333/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":"2503.01333","created_at":"2026-07-05T10:23:04.798157+00:00"},{"alias_kind":"arxiv_version","alias_value":"2503.01333v1","created_at":"2026-07-05T10:23:04.798157+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2503.01333","created_at":"2026-07-05T10:23:04.798157+00:00"},{"alias_kind":"pith_short_12","alias_value":"J47MNN3MCJ56","created_at":"2026-07-05T10:23:04.798157+00:00"},{"alias_kind":"pith_short_16","alias_value":"J47MNN3MCJ56E36H","created_at":"2026-07-05T10:23:04.798157+00:00"},{"alias_kind":"pith_short_8","alias_value":"J47MNN3M","created_at":"2026-07-05T10:23:04.798157+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.05523","citing_title":"CHASE: Adversarial Red-Blue Teaming for Improving LLM Safety using Reinforcement Learning","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2606.28707","citing_title":"BV-Blend: Uncertainty-Weighted Historical Baselines for Stable Critic-Free RL with Verifiable Rewards","ref_index":119,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/J47MNN3MCJ56E36HF4TAA7YY46","json":"https://pith.science/pith/J47MNN3MCJ56E36HF4TAA7YY46.json","graph_json":"https://pith.science/api/pith-number/J47MNN3MCJ56E36HF4TAA7YY46/graph.json","events_json":"https://pith.science/api/pith-number/J47MNN3MCJ56E36HF4TAA7YY46/events.json","paper":"https://pith.science/paper/J47MNN3M"},"agent_actions":{"view_html":"https://pith.science/pith/J47MNN3MCJ56E36HF4TAA7YY46","download_json":"https://pith.science/pith/J47MNN3MCJ56E36HF4TAA7YY46.json","view_paper":"https://pith.science/paper/J47MNN3M","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2503.01333&json=true","fetch_graph":"https://pith.science/api/pith-number/J47MNN3MCJ56E36HF4TAA7YY46/graph.json","fetch_events":"https://pith.science/api/pith-number/J47MNN3MCJ56E36HF4TAA7YY46/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/J47MNN3MCJ56E36HF4TAA7YY46/action/timestamp_anchor","attest_storage":"https://pith.science/pith/J47MNN3MCJ56E36HF4TAA7YY46/action/storage_attestation","attest_author":"https://pith.science/pith/J47MNN3MCJ56E36HF4TAA7YY46/action/author_attestation","sign_citation":"https://pith.science/pith/J47MNN3MCJ56E36HF4TAA7YY46/action/citation_signature","submit_replication":"https://pith.science/pith/J47MNN3MCJ56E36HF4TAA7YY46/action/replication_record"}},"created_at":"2026-07-05T10:23:04.798157+00:00","updated_at":"2026-07-05T10:23:04.798157+00:00"}