{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:JBILANH55GDATESG4NFMNVVH4Z","short_pith_number":"pith:JBILANH5","schema_version":"1.0","canonical_sha256":"4850b034fde986099246e34ac6d6a7e64c961d35d29f0ac269864c56829f1299","source":{"kind":"arxiv","id":"2605.00435","version":2},"attestation_state":"computed","paper":{"title":"Escaping Mode Collapse in LLM Generation via Geometric Regulation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Reinforced Mode Regulation uses low-rank damping in the value cache to prevent geometric collapse and sustain diverse LLM generation at entropy rates down to 0.8 nats per step.","cross_cats":["cond-mat.dis-nn","cs.AI","nlin.CD"],"primary_cat":"cs.CL","authors_text":"Kumiko Tanaka-Ishii, Xin Du","submitted_at":"2026-05-01T06:12:05Z","abstract_excerpt":"Mode collapse is a persistent challenge in generative modeling and appears in autoregressive text generation as behaviors ranging from explicit looping to gradual loss of diversity and premature trajectory convergence. We take a dynamical-systems view and reinterpret mode collapse as reduced state-space accessibility caused by *geometric collapse*: during generation, the model's internal trajectory becomes confined to a low-dimensional region of its representation space. This implies mode collapse is not purely a token-level phenomenon and cannot be reliably solved by symbolic constraints or p"},"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":"2605.00435","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-01T06:12:05Z","cross_cats_sorted":["cond-mat.dis-nn","cs.AI","nlin.CD"],"title_canon_sha256":"79d4d23ee67ac94c33888fad6994faae624a2f901b8bbdfd89ac8e1f173d2a44","abstract_canon_sha256":"8bd2411a8632bd9034bbffdf416de4abef09be42b98fa1cdb0baa9a4f821f1ca"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-28T02:04:48.399227Z","signature_b64":"mOUir9Zy7Qmy4NJUnjIr9hDThonLAIFsi3mxeqdPN6vqnDlIg5bMAjGlc6u8RUWvC+fVGTd8ZsOXTYnYao8vAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4850b034fde986099246e34ac6d6a7e64c961d35d29f0ac269864c56829f1299","last_reissued_at":"2026-05-28T02:04:48.398722Z","signature_status":"signed_v1","first_computed_at":"2026-05-28T02:04:48.398722Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Escaping Mode Collapse in LLM Generation via Geometric Regulation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Reinforced Mode Regulation uses low-rank damping in the value cache to prevent geometric collapse and sustain diverse LLM generation at entropy rates down to 0.8 nats per step.","cross_cats":["cond-mat.dis-nn","cs.AI","nlin.CD"],"primary_cat":"cs.CL","authors_text":"Kumiko Tanaka-Ishii, Xin Du","submitted_at":"2026-05-01T06:12:05Z","abstract_excerpt":"Mode collapse is a persistent challenge in generative modeling and appears in autoregressive text generation as behaviors ranging from explicit looping to gradual loss of diversity and premature trajectory convergence. We take a dynamical-systems view and reinterpret mode collapse as reduced state-space accessibility caused by *geometric collapse*: during generation, the model's internal trajectory becomes confined to a low-dimensional region of its representation space. This implies mode collapse is not purely a token-level phenomenon and cannot be reliably solved by symbolic constraints or p"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"RMR substantially reduces mode collapse and enables stable, high-quality generation at extremely low entropy rates (down to 0.8 nats/step), whereas standard decoding typically collapses near 2.0 nats/step.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That low-rank damping of dominant directions in the value cache directly counters geometric collapse in representation space without introducing new artifacts or quality loss, and that this intervention is sufficient across model scales and tasks.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Reinforced Mode Regulation (RMR) uses low-rank damping on the value cache to prevent geometric collapse and mode collapse in autoregressive LLM generation, supporting stable output down to 0.8 nats/step entropy.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Reinforced Mode Regulation uses low-rank damping in the value cache to prevent geometric collapse and sustain diverse LLM generation at entropy rates down to 0.8 nats per step.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"2fbcbd28fcafb08fbbc8d05e0f557385516324cea39a1fae6e3309b92c3b180f"},"source":{"id":"2605.00435","kind":"arxiv","version":2},"verdict":{"id":"a0e78f54-8f71-4099-bf20-f1027c560d9d","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-09T19:14:44.692056Z","strongest_claim":"RMR substantially reduces mode collapse and enables stable, high-quality generation at extremely low entropy rates (down to 0.8 nats/step), whereas standard decoding typically collapses near 2.0 nats/step.","one_line_summary":"Reinforced Mode Regulation (RMR) uses low-rank damping on the value cache to prevent geometric collapse and mode collapse in autoregressive LLM generation, supporting stable output down to 0.8 nats/step entropy.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That low-rank damping of dominant directions in the value cache directly counters geometric collapse in representation space without introducing new artifacts or quality loss, and that this intervention is sufficient across model scales and tasks.","pith_extraction_headline":"Reinforced Mode Regulation uses low-rank damping in the value cache to prevent geometric collapse and sustain diverse LLM generation at entropy rates down to 0.8 nats per step."},"integrity":{"clean":false,"summary":{"advisory":2,"critical":0,"by_detector":{"doi_compliance":{"total":2,"advisory":2,"critical":0,"informational":0}},"informational":0},"endpoint":"/pith/2605.00435/integrity.json","findings":[{"note":"DOI in the printed bibliography is fragmented by whitespace or line breaks. A longer candidate (10.48550/arXiv.2308.10248.URL) was visible in the surrounding text but could not be confirmed against doi.org as printed.","detector":"doi_compliance","severity":"advisory","ref_index":9,"audited_at":"2026-05-19T18:10:27.103384Z","detected_doi":"10.48550/arXiv.2308.10248.URL","finding_type":"recoverable_identifier","verdict_class":"incontrovertible","detected_arxiv_id":null},{"note":"DOI in the printed bibliography is fragmented by whitespace or line breaks. A longer candidate (10.48550/arXiv.2310.01405.URL) was visible in the surrounding text but could not be confirmed against doi.org as printed.","detector":"doi_compliance","severity":"advisory","ref_index":12,"audited_at":"2026-05-19T18:10:27.103384Z","detected_doi":"10.48550/arXiv.2310.01405.URL","finding_type":"recoverable_identifier","verdict_class":"incontrovertible","detected_arxiv_id":null}],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T19:42:49.248257Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T18:10:27.103384Z","status":"completed","version":"1.0.0","findings_count":2}],"snapshot_sha256":"8c73c8310acdb63eeac948ecbbbef1d101ec2f0b7cd6c2d9cf31f9b004a4437f"},"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":"2605.00435","created_at":"2026-05-28T02:04:48.398789+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.00435v2","created_at":"2026-05-28T02:04:48.398789+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.00435","created_at":"2026-05-28T02:04:48.398789+00:00"},{"alias_kind":"pith_short_12","alias_value":"JBILANH55GDA","created_at":"2026-05-28T02:04:48.398789+00:00"},{"alias_kind":"pith_short_16","alias_value":"JBILANH55GDATESG","created_at":"2026-05-28T02:04:48.398789+00:00"},{"alias_kind":"pith_short_8","alias_value":"JBILANH5","created_at":"2026-05-28T02:04:48.398789+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/JBILANH55GDATESG4NFMNVVH4Z","json":"https://pith.science/pith/JBILANH55GDATESG4NFMNVVH4Z.json","graph_json":"https://pith.science/api/pith-number/JBILANH55GDATESG4NFMNVVH4Z/graph.json","events_json":"https://pith.science/api/pith-number/JBILANH55GDATESG4NFMNVVH4Z/events.json","paper":"https://pith.science/paper/JBILANH5"},"agent_actions":{"view_html":"https://pith.science/pith/JBILANH55GDATESG4NFMNVVH4Z","download_json":"https://pith.science/pith/JBILANH55GDATESG4NFMNVVH4Z.json","view_paper":"https://pith.science/paper/JBILANH5","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.00435&json=true","fetch_graph":"https://pith.science/api/pith-number/JBILANH55GDATESG4NFMNVVH4Z/graph.json","fetch_events":"https://pith.science/api/pith-number/JBILANH55GDATESG4NFMNVVH4Z/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/JBILANH55GDATESG4NFMNVVH4Z/action/timestamp_anchor","attest_storage":"https://pith.science/pith/JBILANH55GDATESG4NFMNVVH4Z/action/storage_attestation","attest_author":"https://pith.science/pith/JBILANH55GDATESG4NFMNVVH4Z/action/author_attestation","sign_citation":"https://pith.science/pith/JBILANH55GDATESG4NFMNVVH4Z/action/citation_signature","submit_replication":"https://pith.science/pith/JBILANH55GDATESG4NFMNVVH4Z/action/replication_record"}},"created_at":"2026-05-28T02:04:48.398789+00:00","updated_at":"2026-05-28T02:04:48.398789+00:00"}