{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:IFB5D6HK3P75ZDVTL7JPX6SZZ4","short_pith_number":"pith:IFB5D6HK","schema_version":"1.0","canonical_sha256":"4143d1f8eadbffdc8eb35fd2fbfa59cf0a3a027b6448027dd0c1062b209d1940","source":{"kind":"arxiv","id":"2506.13502","version":3},"attestation_state":"computed","paper":{"title":"BOW: Training Language Models to Reason Over Plausible Next Words","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Ben Zhou, Jacob Dineen, Ming Shen, Xiao Ye, Zhikun Xu","submitted_at":"2025-06-16T13:58:54Z","abstract_excerpt":"Next-word prediction (NWP) trains language models against a single observed continuation, even though many contexts admit multiple plausible next words. Recent RL-based next-word reasoning methods make this tension explicit: they reward a model for producing a rationale that supports one context-conditioned continuation, which can turn a pre-existing preference into a confident, self-justifying trajectory. We introduce BOW, an RL framework that instead trains models to produce self-contained, neutral, and comprehensive descriptions of the plausible next-word space. BOW's core reward is mediate"},"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":"2506.13502","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2025-06-16T13:58:54Z","cross_cats_sorted":[],"title_canon_sha256":"85ba6929ccb975a16ce284e929b415d74b180f4322a8b3a1258cfbd7182fd093","abstract_canon_sha256":"1c75258834a1a0092e8e292d545fa126dc706de027157e953dde942f36e5e933"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-08-05T01:36:14.302430Z","signature_b64":"/fUBRQkVYxLBDPKoMols0y7HeNLxTu7hs+cHjI17LvE9ySVDtAo0VS7eIkV31gRqui2v/1JwwlxE/9ps1B9dAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"4143d1f8eadbffdc8eb35fd2fbfa59cf0a3a027b6448027dd0c1062b209d1940","last_reissued_at":"2026-08-05T01:36:14.300576Z","signature_status":"signed_v1","first_computed_at":"2026-08-05T01:36:14.300576Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"BOW: Training Language Models to Reason Over Plausible Next Words","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Ben Zhou, Jacob Dineen, Ming Shen, Xiao Ye, Zhikun Xu","submitted_at":"2025-06-16T13:58:54Z","abstract_excerpt":"Next-word prediction (NWP) trains language models against a single observed continuation, even though many contexts admit multiple plausible next words. Recent RL-based next-word reasoning methods make this tension explicit: they reward a model for producing a rationale that supports one context-conditioned continuation, which can turn a pre-existing preference into a confident, self-justifying trajectory. We introduce BOW, an RL framework that instead trains models to produce self-contained, neutral, and comprehensive descriptions of the plausible next-word space. BOW's core reward is mediate"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.13502","kind":"arxiv","version":3},"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/2506.13502/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":"2506.13502","created_at":"2026-08-05T01:36:14.301153+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.13502v3","created_at":"2026-08-05T01:36:14.301153+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.13502","created_at":"2026-08-05T01:36:14.301153+00:00"},{"alias_kind":"pith_short_12","alias_value":"IFB5D6HK3P75","created_at":"2026-08-05T01:36:14.301153+00:00"},{"alias_kind":"pith_short_16","alias_value":"IFB5D6HK3P75ZDVT","created_at":"2026-08-05T01:36:14.301153+00:00"},{"alias_kind":"pith_short_8","alias_value":"IFB5D6HK","created_at":"2026-08-05T01:36:14.301153+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":2,"sample":[{"citing_arxiv_id":"2605.10810","citing_title":"Likelihood scoring for continuations of mathematical text: a self-supervised benchmark with tests for shortcut vulnerabilities","ref_index":10,"is_internal_anchor":true},{"citing_arxiv_id":"2509.10746","citing_title":"RECAP: Transparent Inference-Time Emotion Alignment for Medical Dialogue Systems","ref_index":43,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/IFB5D6HK3P75ZDVTL7JPX6SZZ4","json":"https://pith.science/pith/IFB5D6HK3P75ZDVTL7JPX6SZZ4.json","graph_json":"https://pith.science/api/pith-number/IFB5D6HK3P75ZDVTL7JPX6SZZ4/graph.json","events_json":"https://pith.science/api/pith-number/IFB5D6HK3P75ZDVTL7JPX6SZZ4/events.json","paper":"https://pith.science/paper/IFB5D6HK"},"agent_actions":{"view_html":"https://pith.science/pith/IFB5D6HK3P75ZDVTL7JPX6SZZ4","download_json":"https://pith.science/pith/IFB5D6HK3P75ZDVTL7JPX6SZZ4.json","view_paper":"https://pith.science/paper/IFB5D6HK","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.13502&json=true","fetch_graph":"https://pith.science/api/pith-number/IFB5D6HK3P75ZDVTL7JPX6SZZ4/graph.json","fetch_events":"https://pith.science/api/pith-number/IFB5D6HK3P75ZDVTL7JPX6SZZ4/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/IFB5D6HK3P75ZDVTL7JPX6SZZ4/action/timestamp_anchor","attest_storage":"https://pith.science/pith/IFB5D6HK3P75ZDVTL7JPX6SZZ4/action/storage_attestation","attest_author":"https://pith.science/pith/IFB5D6HK3P75ZDVTL7JPX6SZZ4/action/author_attestation","sign_citation":"https://pith.science/pith/IFB5D6HK3P75ZDVTL7JPX6SZZ4/action/citation_signature","submit_replication":"https://pith.science/pith/IFB5D6HK3P75ZDVTL7JPX6SZZ4/action/replication_record"}},"created_at":"2026-08-05T01:36:14.301153+00:00","updated_at":"2026-08-05T01:36:14.301153+00:00"}