{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:FYM2NAGGVIDTANF6UJBSWR3RNR","short_pith_number":"pith:FYM2NAGG","schema_version":"1.0","canonical_sha256":"2e19a680c6aa073034bea2432b47716c7395fe00e51908cc6a7ffebe999f0b0b","source":{"kind":"arxiv","id":"2502.09650","version":2},"attestation_state":"computed","paper":{"title":"Principled Data Selection for Alignment: The Hidden Risks of Difficult Examples","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Chengqian Gao, Haonan Li, Liu Liu, Peilin Zhao, Zeke Xie, Zhiqiang Xu","submitted_at":"2025-02-11T17:01:11Z","abstract_excerpt":"The alignment of large language models (LLMs) often assumes that using more clean data yields better outcomes, overlooking the match between model capacity and example difficulty. Challenging this, we propose a new principle: Preference data vary in difficulty, and overly difficult examples hinder alignment, by exceeding the model's capacity. Through systematic experimentation, we validate this principle with three key findings: (1) preference examples vary in difficulty, as evidenced by consistent learning orders across alignment runs; (2) overly difficult examples significantly degrade perfo"},"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":"2502.09650","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-02-11T17:01:11Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"f59d679d92f89f77a36ce7e3bb7f3e80c520ef62f0ec84b76deb729a5b99c884","abstract_canon_sha256":"3670c56e4b86004dee33659c496ea243699bb071a7bc765c5d52d136be378aa9"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:02:41.767744Z","signature_b64":"sdUtRlygtdYx5Cp0t+DnypvBdF9Pkk/4DEPtJQ5CKUoEbrac1It4mutuOyJ+gJ7iGte2NuaqJkBZM+qvcbkaAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"2e19a680c6aa073034bea2432b47716c7395fe00e51908cc6a7ffebe999f0b0b","last_reissued_at":"2026-07-05T11:02:41.767228Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:02:41.767228Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Principled Data Selection for Alignment: The Hidden Risks of Difficult Examples","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Chengqian Gao, Haonan Li, Liu Liu, Peilin Zhao, Zeke Xie, Zhiqiang Xu","submitted_at":"2025-02-11T17:01:11Z","abstract_excerpt":"The alignment of large language models (LLMs) often assumes that using more clean data yields better outcomes, overlooking the match between model capacity and example difficulty. Challenging this, we propose a new principle: Preference data vary in difficulty, and overly difficult examples hinder alignment, by exceeding the model's capacity. Through systematic experimentation, we validate this principle with three key findings: (1) preference examples vary in difficulty, as evidenced by consistent learning orders across alignment runs; (2) overly difficult examples significantly degrade perfo"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.09650","kind":"arxiv","version":2},"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/2502.09650/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":"2502.09650","created_at":"2026-07-05T11:02:41.767290+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.09650v2","created_at":"2026-07-05T11:02:41.767290+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.09650","created_at":"2026-07-05T11:02:41.767290+00:00"},{"alias_kind":"pith_short_12","alias_value":"FYM2NAGGVIDT","created_at":"2026-07-05T11:02:41.767290+00:00"},{"alias_kind":"pith_short_16","alias_value":"FYM2NAGGVIDTANF6","created_at":"2026-07-05T11:02:41.767290+00:00"},{"alias_kind":"pith_short_8","alias_value":"FYM2NAGG","created_at":"2026-07-05T11:02:41.767290+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.09078","citing_title":"The Hidden Bias of Process Reward Models:PRISM for Rewarding the Right Reasoning","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2602.07340","citing_title":"Revisiting Robustness for LLM Safety Alignment via Selective Geometry Control","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2508.04149","citing_title":"Difficulty-Based Preference Data Selection by DPO Implicit Reward Gap","ref_index":20,"is_internal_anchor":false},{"citing_arxiv_id":"2604.14314","citing_title":"DharmaOCR: Specialized Small Language Models for Structured OCR that outperform Open-Source and Commercial Baselines","ref_index":54,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/FYM2NAGGVIDTANF6UJBSWR3RNR","json":"https://pith.science/pith/FYM2NAGGVIDTANF6UJBSWR3RNR.json","graph_json":"https://pith.science/api/pith-number/FYM2NAGGVIDTANF6UJBSWR3RNR/graph.json","events_json":"https://pith.science/api/pith-number/FYM2NAGGVIDTANF6UJBSWR3RNR/events.json","paper":"https://pith.science/paper/FYM2NAGG"},"agent_actions":{"view_html":"https://pith.science/pith/FYM2NAGGVIDTANF6UJBSWR3RNR","download_json":"https://pith.science/pith/FYM2NAGGVIDTANF6UJBSWR3RNR.json","view_paper":"https://pith.science/paper/FYM2NAGG","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.09650&json=true","fetch_graph":"https://pith.science/api/pith-number/FYM2NAGGVIDTANF6UJBSWR3RNR/graph.json","fetch_events":"https://pith.science/api/pith-number/FYM2NAGGVIDTANF6UJBSWR3RNR/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/FYM2NAGGVIDTANF6UJBSWR3RNR/action/timestamp_anchor","attest_storage":"https://pith.science/pith/FYM2NAGGVIDTANF6UJBSWR3RNR/action/storage_attestation","attest_author":"https://pith.science/pith/FYM2NAGGVIDTANF6UJBSWR3RNR/action/author_attestation","sign_citation":"https://pith.science/pith/FYM2NAGGVIDTANF6UJBSWR3RNR/action/citation_signature","submit_replication":"https://pith.science/pith/FYM2NAGGVIDTANF6UJBSWR3RNR/action/replication_record"}},"created_at":"2026-07-05T11:02:41.767290+00:00","updated_at":"2026-07-05T11:02:41.767290+00:00"}