{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:CUMTZLPDYJSPZ2KA3RNR2MKCMO","short_pith_number":"pith:CUMTZLPD","schema_version":"1.0","canonical_sha256":"15193cade3c264fce940dc5b1d3142638807ec413e4da723d8734684ca18716f","source":{"kind":"arxiv","id":"2505.13988","version":1},"attestation_state":"computed","paper":{"title":"The Hallucination Tax of Reinforcement Finetuning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Jieyu Zhao, Linxin Song, Taiwei Shi","submitted_at":"2025-05-20T06:36:45Z","abstract_excerpt":"Reinforcement finetuning (RFT) has become a standard approach for enhancing the reasoning capabilities of large language models (LLMs). However, its impact on model trustworthiness remains underexplored. In this work, we identify and systematically study a critical side effect of RFT, which we term the hallucination tax: a degradation in refusal behavior causing models to produce hallucinated answers to unanswerable questions confidently. To investigate this, we introduce SUM (Synthetic Unanswerable Math), a high-quality dataset of unanswerable math problems designed to probe models' ability t"},"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":"2505.13988","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-05-20T06:36:45Z","cross_cats_sorted":[],"title_canon_sha256":"bf2be439dfcdaa0467c07dd383749a349971bf427c7fcb11e3f0b5adb6fff762","abstract_canon_sha256":"8a7b2c19cd1faea19eb8d0c12bdd7f7ad6ef0a3a26808835d040eb55c39d188c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:05:54.327883Z","signature_b64":"lS9nf6qFVX4JYMojZkq6FrPYX8JjoGK9VlVBIFRwqZTq8rPNrwcFHQtkzQtgZnDMnqtUQtEjbFSYK47ysC7TCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"15193cade3c264fce940dc5b1d3142638807ec413e4da723d8734684ca18716f","last_reissued_at":"2026-07-05T11:05:54.327449Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:05:54.327449Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"The Hallucination Tax of Reinforcement Finetuning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Jieyu Zhao, Linxin Song, Taiwei Shi","submitted_at":"2025-05-20T06:36:45Z","abstract_excerpt":"Reinforcement finetuning (RFT) has become a standard approach for enhancing the reasoning capabilities of large language models (LLMs). However, its impact on model trustworthiness remains underexplored. In this work, we identify and systematically study a critical side effect of RFT, which we term the hallucination tax: a degradation in refusal behavior causing models to produce hallucinated answers to unanswerable questions confidently. To investigate this, we introduce SUM (Synthetic Unanswerable Math), a high-quality dataset of unanswerable math problems designed to probe models' ability t"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.13988","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/2505.13988/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":"2505.13988","created_at":"2026-07-05T11:05:54.327502+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.13988v1","created_at":"2026-07-05T11:05:54.327502+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.13988","created_at":"2026-07-05T11:05:54.327502+00:00"},{"alias_kind":"pith_short_12","alias_value":"CUMTZLPDYJSP","created_at":"2026-07-05T11:05:54.327502+00:00"},{"alias_kind":"pith_short_16","alias_value":"CUMTZLPDYJSPZ2KA","created_at":"2026-07-05T11:05:54.327502+00:00"},{"alias_kind":"pith_short_8","alias_value":"CUMTZLPD","created_at":"2026-07-05T11:05:54.327502+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.07812","citing_title":"Scaling Participation in Modular AI Systems","ref_index":81,"is_internal_anchor":false},{"citing_arxiv_id":"2605.28070","citing_title":"Bridging the Detection-to-Abstention Gap in Reasoning Models under Insufficient Information","ref_index":42,"is_internal_anchor":false},{"citing_arxiv_id":"2604.27859","citing_title":"Rethinking Agentic Reinforcement Learning In Large Language Models","ref_index":79,"is_internal_anchor":false},{"citing_arxiv_id":"2601.21257","citing_title":"MoCo: A One-Stop Shop for Model Collaboration Research","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2604.27859","citing_title":"Rethinking Agentic Reinforcement Learning In Large Language Models","ref_index":79,"is_internal_anchor":false},{"citing_arxiv_id":"2604.27859","citing_title":"Rethinking Agentic Reinforcement Learning In Large Language Models","ref_index":79,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/CUMTZLPDYJSPZ2KA3RNR2MKCMO","json":"https://pith.science/pith/CUMTZLPDYJSPZ2KA3RNR2MKCMO.json","graph_json":"https://pith.science/api/pith-number/CUMTZLPDYJSPZ2KA3RNR2MKCMO/graph.json","events_json":"https://pith.science/api/pith-number/CUMTZLPDYJSPZ2KA3RNR2MKCMO/events.json","paper":"https://pith.science/paper/CUMTZLPD"},"agent_actions":{"view_html":"https://pith.science/pith/CUMTZLPDYJSPZ2KA3RNR2MKCMO","download_json":"https://pith.science/pith/CUMTZLPDYJSPZ2KA3RNR2MKCMO.json","view_paper":"https://pith.science/paper/CUMTZLPD","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.13988&json=true","fetch_graph":"https://pith.science/api/pith-number/CUMTZLPDYJSPZ2KA3RNR2MKCMO/graph.json","fetch_events":"https://pith.science/api/pith-number/CUMTZLPDYJSPZ2KA3RNR2MKCMO/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/CUMTZLPDYJSPZ2KA3RNR2MKCMO/action/timestamp_anchor","attest_storage":"https://pith.science/pith/CUMTZLPDYJSPZ2KA3RNR2MKCMO/action/storage_attestation","attest_author":"https://pith.science/pith/CUMTZLPDYJSPZ2KA3RNR2MKCMO/action/author_attestation","sign_citation":"https://pith.science/pith/CUMTZLPDYJSPZ2KA3RNR2MKCMO/action/citation_signature","submit_replication":"https://pith.science/pith/CUMTZLPDYJSPZ2KA3RNR2MKCMO/action/replication_record"}},"created_at":"2026-07-05T11:05:54.327502+00:00","updated_at":"2026-07-05T11:05:54.327502+00:00"}