{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:LKDC5EVBTAIZWCBIBJTBDKSAQ2","short_pith_number":"pith:LKDC5EVB","schema_version":"1.0","canonical_sha256":"5a862e92a198119b08280a6611aa4086a55c451ca844be3fff317c1cfdec73e6","source":{"kind":"arxiv","id":"2505.18830","version":1},"attestation_state":"computed","paper":{"title":"On the Effect of Negative Gradient in Group Relative Deep Reinforcement Optimization","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Christos Thrampoulidis, Danica J. Sutherland, Muchen Li, Wenlong Deng, Xiaoxiao Li, Yi Ren","submitted_at":"2025-05-24T18:58:51Z","abstract_excerpt":"Reinforcement learning (RL) has become popular in enhancing the reasoning capabilities of large language models (LLMs), with Group Relative Policy Optimization (GRPO) emerging as a widely used algorithm in recent systems. Despite GRPO's widespread adoption, we identify a previously unrecognized phenomenon we term Lazy Likelihood Displacement (LLD), wherein the likelihood of correct responses marginally increases or even decreases during training. This behavior mirrors a recently discovered misalignment issue in Direct Preference Optimization (DPO), attributed to the influence of negative gradi"},"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":"2505.18830","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-05-24T18:58:51Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"3ab46ee6207538fac2627bd27c6b2c140dbd52298f6aa7ee00b7a3b0c47193ad","abstract_canon_sha256":"38a9c0a132d78f5fb18aadccf9d72277b6d5258a12ca48f61ef67b673767111e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:09:20.213942Z","signature_b64":"Y04rI3OAWB6kMLwE2kdmyHqLGSSYHl7h3iDGqZURUIOUuGIwgt0u0JS1w8vwe6Y9HjIH4wYe44KeAAudVdDICg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"5a862e92a198119b08280a6611aa4086a55c451ca844be3fff317c1cfdec73e6","last_reissued_at":"2026-07-05T11:09:20.213467Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:09:20.213467Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"On the Effect of Negative Gradient in Group Relative Deep Reinforcement Optimization","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Christos Thrampoulidis, Danica J. Sutherland, Muchen Li, Wenlong Deng, Xiaoxiao Li, Yi Ren","submitted_at":"2025-05-24T18:58:51Z","abstract_excerpt":"Reinforcement learning (RL) has become popular in enhancing the reasoning capabilities of large language models (LLMs), with Group Relative Policy Optimization (GRPO) emerging as a widely used algorithm in recent systems. Despite GRPO's widespread adoption, we identify a previously unrecognized phenomenon we term Lazy Likelihood Displacement (LLD), wherein the likelihood of correct responses marginally increases or even decreases during training. This behavior mirrors a recently discovered misalignment issue in Direct Preference Optimization (DPO), attributed to the influence of negative gradi"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.18830","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/2505.18830/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":"2505.18830","created_at":"2026-07-05T11:09:20.213524+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.18830v1","created_at":"2026-07-05T11:09:20.213524+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.18830","created_at":"2026-07-05T11:09:20.213524+00:00"},{"alias_kind":"pith_short_12","alias_value":"LKDC5EVBTAIZ","created_at":"2026-07-05T11:09:20.213524+00:00"},{"alias_kind":"pith_short_16","alias_value":"LKDC5EVBTAIZWCBI","created_at":"2026-07-05T11:09:20.213524+00:00"},{"alias_kind":"pith_short_8","alias_value":"LKDC5EVB","created_at":"2026-07-05T11:09:20.213524+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.00347","citing_title":"Odysseus: Scaling VLMs to 100+ Turn Decision-Making in Games via Reinforcement Learning","ref_index":78,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/LKDC5EVBTAIZWCBIBJTBDKSAQ2","json":"https://pith.science/pith/LKDC5EVBTAIZWCBIBJTBDKSAQ2.json","graph_json":"https://pith.science/api/pith-number/LKDC5EVBTAIZWCBIBJTBDKSAQ2/graph.json","events_json":"https://pith.science/api/pith-number/LKDC5EVBTAIZWCBIBJTBDKSAQ2/events.json","paper":"https://pith.science/paper/LKDC5EVB"},"agent_actions":{"view_html":"https://pith.science/pith/LKDC5EVBTAIZWCBIBJTBDKSAQ2","download_json":"https://pith.science/pith/LKDC5EVBTAIZWCBIBJTBDKSAQ2.json","view_paper":"https://pith.science/paper/LKDC5EVB","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.18830&json=true","fetch_graph":"https://pith.science/api/pith-number/LKDC5EVBTAIZWCBIBJTBDKSAQ2/graph.json","fetch_events":"https://pith.science/api/pith-number/LKDC5EVBTAIZWCBIBJTBDKSAQ2/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/LKDC5EVBTAIZWCBIBJTBDKSAQ2/action/timestamp_anchor","attest_storage":"https://pith.science/pith/LKDC5EVBTAIZWCBIBJTBDKSAQ2/action/storage_attestation","attest_author":"https://pith.science/pith/LKDC5EVBTAIZWCBIBJTBDKSAQ2/action/author_attestation","sign_citation":"https://pith.science/pith/LKDC5EVBTAIZWCBIBJTBDKSAQ2/action/citation_signature","submit_replication":"https://pith.science/pith/LKDC5EVBTAIZWCBIBJTBDKSAQ2/action/replication_record"}},"created_at":"2026-07-05T11:09:20.213524+00:00","updated_at":"2026-07-05T11:09:20.213524+00:00"}