{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:Z7QLAXPZNEOC6HXYDPQ6MFIUMY","short_pith_number":"pith:Z7QLAXPZ","schema_version":"1.0","canonical_sha256":"cfe0b05df9691c2f1ef81be1e615146610fa8b61949ed71f6c188e9c19d90c27","source":{"kind":"arxiv","id":"2605.14366","version":1},"attestation_state":"computed","paper":{"title":"Reinforcement Learning with Semantic Rewards Enables Low-Resource Language Expansion without Alignment Tax","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Reinforcement learning with embedding-level semantic rewards lets LLMs add low-resource languages without the usual loss of general skills.","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Guixian Xu, Longfei Zheng, Rong Fu, Wentao Zhang, Xiaolu Zhang, Xuexian Song, Zeli Su, Zhankai Xu, Zhou Liu, Ziyin Zhang","submitted_at":"2026-05-14T04:47:22Z","abstract_excerpt":"Extending large language models (LLMs) to low-resource languages often incurs an \"alignment tax\": improvements in the target language come at the cost of catastrophic forgetting in general capabilities. We argue that this trade-off arises from the rigidity of supervised fine-tuning (SFT), which enforces token-level surface imitation on narrow and biased data distributions. To address this limitation, we propose a semantic-space alignment paradigm powered by Group Relative Policy Optimization (GRPO), where the model is optimized using embedding-level semantic rewards rather than likelihood maxi"},"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":true,"formal_links_present":true},"canonical_record":{"source":{"id":"2605.14366","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-14T04:47:22Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"f4863bbe0621727b76fdaf7b4479f8558cd16a2da763d0125b128359c6c033ea","abstract_canon_sha256":"49333ace483f564e023ef1d6138125147e77396eb457ea09bcd626547a90e4d7"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:39:07.886965Z","signature_b64":"r7YM5vvxqpCqPbMR1T4jorgakJF0kbzr8ktFGr44ZuCNKivmPVPTCjiAEtJe1BbboiPsNmNoz6JQcyxeMelNAg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"cfe0b05df9691c2f1ef81be1e615146610fa8b61949ed71f6c188e9c19d90c27","last_reissued_at":"2026-05-17T23:39:07.886359Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:39:07.886359Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Reinforcement Learning with Semantic Rewards Enables Low-Resource Language Expansion without Alignment Tax","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Reinforcement learning with embedding-level semantic rewards lets LLMs add low-resource languages without the usual loss of general skills.","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Guixian Xu, Longfei Zheng, Rong Fu, Wentao Zhang, Xiaolu Zhang, Xuexian Song, Zeli Su, Zhankai Xu, Zhou Liu, Ziyin Zhang","submitted_at":"2026-05-14T04:47:22Z","abstract_excerpt":"Extending large language models (LLMs) to low-resource languages often incurs an \"alignment tax\": improvements in the target language come at the cost of catastrophic forgetting in general capabilities. We argue that this trade-off arises from the rigidity of supervised fine-tuning (SFT), which enforces token-level surface imitation on narrow and biased data distributions. To address this limitation, we propose a semantic-space alignment paradigm powered by Group Relative Policy Optimization (GRPO), where the model is optimized using embedding-level semantic rewards rather than likelihood maxi"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Experiments show that our method acquires low-resource capabilities while markedly mitigating alignment tax, preserving general competence more effectively than SFT.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That embedding-level semantic rewards reliably capture and preserve intended meaning across languages without introducing new biases or requiring the model to have strong pretrained semantic understanding in the target language.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Reinforcement learning with semantic rewards lets LLMs gain low-resource language skills without the alignment tax that degrades general capabilities in supervised fine-tuning.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Reinforcement learning with embedding-level semantic rewards lets LLMs add low-resource languages without the usual loss of general skills.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"5eb63a024a0766558abbc7473f17968c01ede3936cc6cc8a9ceb6dcdae7b5ca8"},"source":{"id":"2605.14366","kind":"arxiv","version":1},"verdict":{"id":"45384cd8-39e2-4bbc-8359-2e5f3f40bf2d","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T02:37:54.778015Z","strongest_claim":"Experiments show that our method acquires low-resource capabilities while markedly mitigating alignment tax, preserving general competence more effectively than SFT.","one_line_summary":"Reinforcement learning with semantic rewards lets LLMs gain low-resource language skills without the alignment tax that degrades general capabilities in supervised fine-tuning.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That embedding-level semantic rewards reliably capture and preserve intended meaning across languages without introducing new biases or requiring the model to have strong pretrained semantic understanding in the target language.","pith_extraction_headline":"Reinforcement learning with embedding-level semantic rewards lets LLMs add low-resource languages without the usual loss of general skills."},"references":{"count":38,"sample":[{"doi":"10.1162/tacl_a_00343","year":2020,"title":"Transactions of the Association for Computational Linguistics , volume =","work_id":"ac8e5e49-0c81-4708-a831-537e1c6797bb","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.18653/v1/2020.coling-main.574","year":2020,"title":"Proceedings of the 28th International Conference on Computational Linguistics , year =","work_id":"68bee7c2-72ce-4164-bf37-371e1b63bada","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.18653/v1/2020.coling-main.381","year":2020,"title":"Proceedings of the 28th International Conference on Computational Linguistics , year =","work_id":"cd404891-c6fb-4234-94e5-fcda67cb5aa8","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"Proceedings of the National Academy of Sciences , volume =","work_id":"7d346159-1571-4ab3-8bda-afb6b1c99afe","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Proceedings of the European Conference on Computer Vision (ECCV) , year =","work_id":"eea8d58b-3b2a-46f6-9c61-070a4414f5d0","ref_index":6,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":38,"snapshot_sha256":"8c26adc7dd01caca208f7f0cf65486be46d7dd0fa7179bfab2222cb57f1644fb","internal_anchors":10},"formal_canon":{"evidence_count":2,"snapshot_sha256":"24e7b85c474bd0a15e68827203481b908cedb3c90af88a1c4575d926fcfc9849"},"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":"2605.14366","created_at":"2026-05-17T23:39:07.886473+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.14366v1","created_at":"2026-05-17T23:39:07.886473+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.14366","created_at":"2026-05-17T23:39:07.886473+00:00"},{"alias_kind":"pith_short_12","alias_value":"Z7QLAXPZNEOC","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"Z7QLAXPZNEOC6HXY","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"Z7QLAXPZ","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/Z7QLAXPZNEOC6HXYDPQ6MFIUMY","json":"https://pith.science/pith/Z7QLAXPZNEOC6HXYDPQ6MFIUMY.json","graph_json":"https://pith.science/api/pith-number/Z7QLAXPZNEOC6HXYDPQ6MFIUMY/graph.json","events_json":"https://pith.science/api/pith-number/Z7QLAXPZNEOC6HXYDPQ6MFIUMY/events.json","paper":"https://pith.science/paper/Z7QLAXPZ"},"agent_actions":{"view_html":"https://pith.science/pith/Z7QLAXPZNEOC6HXYDPQ6MFIUMY","download_json":"https://pith.science/pith/Z7QLAXPZNEOC6HXYDPQ6MFIUMY.json","view_paper":"https://pith.science/paper/Z7QLAXPZ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.14366&json=true","fetch_graph":"https://pith.science/api/pith-number/Z7QLAXPZNEOC6HXYDPQ6MFIUMY/graph.json","fetch_events":"https://pith.science/api/pith-number/Z7QLAXPZNEOC6HXYDPQ6MFIUMY/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/Z7QLAXPZNEOC6HXYDPQ6MFIUMY/action/timestamp_anchor","attest_storage":"https://pith.science/pith/Z7QLAXPZNEOC6HXYDPQ6MFIUMY/action/storage_attestation","attest_author":"https://pith.science/pith/Z7QLAXPZNEOC6HXYDPQ6MFIUMY/action/author_attestation","sign_citation":"https://pith.science/pith/Z7QLAXPZNEOC6HXYDPQ6MFIUMY/action/citation_signature","submit_replication":"https://pith.science/pith/Z7QLAXPZNEOC6HXYDPQ6MFIUMY/action/replication_record"}},"created_at":"2026-05-17T23:39:07.886473+00:00","updated_at":"2026-05-17T23:39:07.886473+00:00"}