{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:4N5DHTUHPZXD6W3UZHQN3KTFQZ","short_pith_number":"pith:4N5DHTUH","schema_version":"1.0","canonical_sha256":"e37a33ce877e6e3f5b74c9e0ddaa6586431feb5f37827b626f6e4adc829db52a","source":{"kind":"arxiv","id":"2505.18949","version":1},"attestation_state":"computed","paper":{"title":"The Price of Format: Diversity Collapse in LLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Chenyang An, Jingbo Shang, Letian Peng, Longfei Yun, Zilong Wang","submitted_at":"2025-05-25T02:52:35Z","abstract_excerpt":"Instruction-tuned large language models (LLMs) employ structured templates, such as role markers and special tokens, to enforce format consistency during inference. However, we identify a critical limitation of such formatting: it induces a phenomenon we term diversity collapse, where the model generates semantically similar outputs for open-ended inputs, undermining creativity and variability. We systematically evaluate this effect across tasks like story completion and free-form generation, finding that (1) diversity collapse persists even under high-temperature sampling, and (2) structural "},"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":"2505.18949","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-05-25T02:52:35Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"a9df32aab10d3746e0aa65eae428604ac15b1b8f57c462036422bfaa087ac41e","abstract_canon_sha256":"ffbb766816df2fc0b28dfaa56f1e359953891651d6837b91a6331ad8102ed2ca"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:09:22.480797Z","signature_b64":"PT7ZnbOOKp3Kg/HMO5e0rhiRel/eYkaxATllDpvvQlAtSJT7aqrhuaeMdzZ5u+QFhsYxKEJk8vmf3nPBZ/nCDw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"e37a33ce877e6e3f5b74c9e0ddaa6586431feb5f37827b626f6e4adc829db52a","last_reissued_at":"2026-07-05T11:09:22.480219Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:09:22.480219Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"The Price of Format: Diversity Collapse in LLMs","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Chenyang An, Jingbo Shang, Letian Peng, Longfei Yun, Zilong Wang","submitted_at":"2025-05-25T02:52:35Z","abstract_excerpt":"Instruction-tuned large language models (LLMs) employ structured templates, such as role markers and special tokens, to enforce format consistency during inference. However, we identify a critical limitation of such formatting: it induces a phenomenon we term diversity collapse, where the model generates semantically similar outputs for open-ended inputs, undermining creativity and variability. We systematically evaluate this effect across tasks like story completion and free-form generation, finding that (1) diversity collapse persists even under high-temperature sampling, and (2) structural "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.18949","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/2505.18949/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":"2505.18949","created_at":"2026-07-05T11:09:22.480279+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.18949v1","created_at":"2026-07-05T11:09:22.480279+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.18949","created_at":"2026-07-05T11:09:22.480279+00:00"},{"alias_kind":"pith_short_12","alias_value":"4N5DHTUHPZXD","created_at":"2026-07-05T11:09:22.480279+00:00"},{"alias_kind":"pith_short_16","alias_value":"4N5DHTUHPZXD6W3U","created_at":"2026-07-05T11:09:22.480279+00:00"},{"alias_kind":"pith_short_8","alias_value":"4N5DHTUH","created_at":"2026-07-05T11:09:22.480279+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":7,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2605.27268","citing_title":"Lost in Sampling: Assessing Lexical Reachability in LLMs via the Word Coverage Score (WCS)","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11258","citing_title":"Unlocking LLM Creativity in Science through Analogical Reasoning","ref_index":55,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08737","citing_title":"The Extrapolation Cliff in On-Policy Distillation of Near-Deterministic Structured Outputs","ref_index":44,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09995","citing_title":"Annotations Mitigate Post-Training Mode Collapse","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06540","citing_title":"Ex Ante Evaluation of AI-Induced Idea Diversity Collapse","ref_index":53,"is_internal_anchor":false},{"citing_arxiv_id":"2604.12223","citing_title":"LLM-Guided Semantic Bootstrapping for Interpretable Text Classification with Tsetlin Machines","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2604.14862","citing_title":"Schema Key Wording as an Instruction Channel in Structured Generation under Constrained Decoding","ref_index":23,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/4N5DHTUHPZXD6W3UZHQN3KTFQZ","json":"https://pith.science/pith/4N5DHTUHPZXD6W3UZHQN3KTFQZ.json","graph_json":"https://pith.science/api/pith-number/4N5DHTUHPZXD6W3UZHQN3KTFQZ/graph.json","events_json":"https://pith.science/api/pith-number/4N5DHTUHPZXD6W3UZHQN3KTFQZ/events.json","paper":"https://pith.science/paper/4N5DHTUH"},"agent_actions":{"view_html":"https://pith.science/pith/4N5DHTUHPZXD6W3UZHQN3KTFQZ","download_json":"https://pith.science/pith/4N5DHTUHPZXD6W3UZHQN3KTFQZ.json","view_paper":"https://pith.science/paper/4N5DHTUH","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.18949&json=true","fetch_graph":"https://pith.science/api/pith-number/4N5DHTUHPZXD6W3UZHQN3KTFQZ/graph.json","fetch_events":"https://pith.science/api/pith-number/4N5DHTUHPZXD6W3UZHQN3KTFQZ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/4N5DHTUHPZXD6W3UZHQN3KTFQZ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/4N5DHTUHPZXD6W3UZHQN3KTFQZ/action/storage_attestation","attest_author":"https://pith.science/pith/4N5DHTUHPZXD6W3UZHQN3KTFQZ/action/author_attestation","sign_citation":"https://pith.science/pith/4N5DHTUHPZXD6W3UZHQN3KTFQZ/action/citation_signature","submit_replication":"https://pith.science/pith/4N5DHTUHPZXD6W3UZHQN3KTFQZ/action/replication_record"}},"created_at":"2026-07-05T11:09:22.480279+00:00","updated_at":"2026-07-05T11:09:22.480279+00:00"}