{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:UHXFDEJWIC2UDGHU4WG4TOBCL2","short_pith_number":"pith:UHXFDEJW","schema_version":"1.0","canonical_sha256":"a1ee51913640b54198f4e58dc9b8225ebc9cc008f8a78a4579883a3cfeae8003","source":{"kind":"arxiv","id":"2605.09781","version":2},"attestation_state":"computed","paper":{"title":"Parameter-Efficient Neuroevolution for Diverse LLM Generation: Quality-Diversity Optimization via Prompt Embedding Evolution","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Evolving compact prompt embeddings inside frozen large language models produces more diverse outputs than standard methods.","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.NE","authors_text":"Dongxin Guo, Jikun Wu, Siu Ming Yiu","submitted_at":"2026-05-10T22:00:15Z","abstract_excerpt":"Large Language Models exhibit mode collapse, producing homogeneous outputs that fail to explore valid solution spaces. We present QD-LLM, a framework for parameter-efficient neuroevolution that evolves prompt embeddings, compact neural interfaces (~32K parameters) that steer generation in frozen LLMs (70B+ parameters), within a Quality-Diversity (QD) optimization framework. Our contributions: (1) evolved prompt embeddings via gradient-free optimization enabling behavioral steering without model fine-tuning; (2) hybrid behavior characterization combining semantic and explicit features with form"},"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":"2605.09781","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.NE","submitted_at":"2026-05-10T22:00:15Z","cross_cats_sorted":["cs.AI","cs.CL","cs.LG"],"title_canon_sha256":"e9dc7ee6fa6dcd19345458b189f00241e0bab47d5c04212ea8c44fb2f527aa93","abstract_canon_sha256":"217074b293386fa91854a82f0b7692a482d477bf22e4e3dde7079bd30490fa69"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-23T01:12:08.168768Z","signature_b64":"FJN/6/OD6bij1c8FGXL3huBC0JY3XEkNNbE5wo2rZJZwO6qDHp7uz0c+dTL6BXXHdiZjlrFIBUVSBtpWCcA3Cg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a1ee51913640b54198f4e58dc9b8225ebc9cc008f8a78a4579883a3cfeae8003","last_reissued_at":"2026-06-23T01:12:08.168300Z","signature_status":"signed_v1","first_computed_at":"2026-06-23T01:12:08.168300Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Parameter-Efficient Neuroevolution for Diverse LLM Generation: Quality-Diversity Optimization via Prompt Embedding Evolution","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Evolving compact prompt embeddings inside frozen large language models produces more diverse outputs than standard methods.","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.NE","authors_text":"Dongxin Guo, Jikun Wu, Siu Ming Yiu","submitted_at":"2026-05-10T22:00:15Z","abstract_excerpt":"Large Language Models exhibit mode collapse, producing homogeneous outputs that fail to explore valid solution spaces. We present QD-LLM, a framework for parameter-efficient neuroevolution that evolves prompt embeddings, compact neural interfaces (~32K parameters) that steer generation in frozen LLMs (70B+ parameters), within a Quality-Diversity (QD) optimization framework. Our contributions: (1) evolved prompt embeddings via gradient-free optimization enabling behavioral steering without model fine-tuning; (2) hybrid behavior characterization combining semantic and explicit features with form"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On HumanEval (164 problems), MBPP, and creative writing benchmarks, QD-LLM achieves 46.4% higher coverage and 41.4% higher QD-Score than QDAIF (p<0.001, 30 runs, Vargha-Delaney A=0.94).","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That hybrid semantic-plus-explicit behavior descriptors remain sufficiently independent (NMI = 0.08 ± 0.02) to support the formal coverage bounds of Theorem 1 and that prompt embeddings of ~32K parameters can reliably steer 70B+ frozen LLMs across the tested domains.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"QD-LLM evolves prompt embeddings via neuroevolution in a quality-diversity framework, delivering 46% higher coverage and 41% higher QD-score than prior methods on coding and writing benchmarks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Evolving compact prompt embeddings inside frozen large language models produces more diverse outputs than standard methods.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"c27d8d8b4888a348b357dd26afcfc826d8fabcaa81ad31373eb4045012101b5a"},"source":{"id":"2605.09781","kind":"arxiv","version":2},"verdict":{"id":"8365227d-ebc1-43e6-882b-f1a5b1d573cc","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-12T03:34:01.371350Z","strongest_claim":"On HumanEval (164 problems), MBPP, and creative writing benchmarks, QD-LLM achieves 46.4% higher coverage and 41.4% higher QD-Score than QDAIF (p<0.001, 30 runs, Vargha-Delaney A=0.94).","one_line_summary":"QD-LLM evolves prompt embeddings via neuroevolution in a quality-diversity framework, delivering 46% higher coverage and 41% higher QD-score than prior methods on coding and writing benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That hybrid semantic-plus-explicit behavior descriptors remain sufficiently independent (NMI = 0.08 ± 0.02) to support the formal coverage bounds of Theorem 1 and that prompt embeddings of ~32K parameters can reliably steer 70B+ frozen LLMs across the tested domains.","pith_extraction_headline":"Evolving compact prompt embeddings inside frozen large language models produces more diverse outputs than standard methods."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.09781/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-20T07:02:01.406459Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T16:36:27.365646Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T12:31:17.892405Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T09:58:13.393502Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"03f6d1de9ff5e7e33a870aea5a7e7e1cc87ea1bbdbee5713065bd8eaef1b9f54"},"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":"2605.09781","created_at":"2026-06-23T01:12:08.168380+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.09781v2","created_at":"2026-06-23T01:12:08.168380+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.09781","created_at":"2026-06-23T01:12:08.168380+00:00"},{"alias_kind":"pith_short_12","alias_value":"UHXFDEJWIC2U","created_at":"2026-06-23T01:12:08.168380+00:00"},{"alias_kind":"pith_short_16","alias_value":"UHXFDEJWIC2UDGHU","created_at":"2026-06-23T01:12:08.168380+00:00"},{"alias_kind":"pith_short_8","alias_value":"UHXFDEJW","created_at":"2026-06-23T01:12:08.168380+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/UHXFDEJWIC2UDGHU4WG4TOBCL2","json":"https://pith.science/pith/UHXFDEJWIC2UDGHU4WG4TOBCL2.json","graph_json":"https://pith.science/api/pith-number/UHXFDEJWIC2UDGHU4WG4TOBCL2/graph.json","events_json":"https://pith.science/api/pith-number/UHXFDEJWIC2UDGHU4WG4TOBCL2/events.json","paper":"https://pith.science/paper/UHXFDEJW"},"agent_actions":{"view_html":"https://pith.science/pith/UHXFDEJWIC2UDGHU4WG4TOBCL2","download_json":"https://pith.science/pith/UHXFDEJWIC2UDGHU4WG4TOBCL2.json","view_paper":"https://pith.science/paper/UHXFDEJW","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.09781&json=true","fetch_graph":"https://pith.science/api/pith-number/UHXFDEJWIC2UDGHU4WG4TOBCL2/graph.json","fetch_events":"https://pith.science/api/pith-number/UHXFDEJWIC2UDGHU4WG4TOBCL2/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/UHXFDEJWIC2UDGHU4WG4TOBCL2/action/timestamp_anchor","attest_storage":"https://pith.science/pith/UHXFDEJWIC2UDGHU4WG4TOBCL2/action/storage_attestation","attest_author":"https://pith.science/pith/UHXFDEJWIC2UDGHU4WG4TOBCL2/action/author_attestation","sign_citation":"https://pith.science/pith/UHXFDEJWIC2UDGHU4WG4TOBCL2/action/citation_signature","submit_replication":"https://pith.science/pith/UHXFDEJWIC2UDGHU4WG4TOBCL2/action/replication_record"}},"created_at":"2026-06-23T01:12:08.168380+00:00","updated_at":"2026-06-23T01:12:08.168380+00:00"}