{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:W7J5BAKA7ZMBEENX4LM3KRP6IH","short_pith_number":"pith:W7J5BAKA","schema_version":"1.0","canonical_sha256":"b7d3d08140fe581211b7e2d9b545fe41c59d121481d24accc851ef8ffaaeb3d2","source":{"kind":"arxiv","id":"2502.06533","version":1},"attestation_state":"computed","paper":{"title":"Ignore the KL Penalty! Boosting Exploration on Critical Tokens to Enhance RL Fine-Tuning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Jean Vassoyan, Nathana\\\"el Beau, Roman Plaud","submitted_at":"2025-02-10T14:56:25Z","abstract_excerpt":"The ability to achieve long-term goals is a key challenge in the current development of large language models (LLMs). To address this, pre-trained LLMs can be fine-tuned with reinforcement learning (RL) to explore solutions that optimize a given goal. However, exploration with LLMs is difficult, as a balance has to be struck between discovering new solutions and staying close enough to the pre-trained model, so as not to degrade basic capabilities. This is typically controlled with a Kullback-Leibler (KL) penalty. In this paper, we investigate the exploration dynamics of a small language model"},"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":"2502.06533","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-02-10T14:56:25Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"fbc7a6a6fd3c407d3362244298d538e1937265f14276285b99c8e10110c8d638","abstract_canon_sha256":"f6ab39ca04b893e16e46f21dbd82cda17cda4289b6e6692009f7eabdabe665f8"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:12:08.117365Z","signature_b64":"1/deQA9qdoLvjAni+EscVerciGGnRs371PJgR+nkJ7ZqssV1S/OrSZZ5A1oxTOjSt1hkf5w2uzf7LnJrrZhtAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b7d3d08140fe581211b7e2d9b545fe41c59d121481d24accc851ef8ffaaeb3d2","last_reissued_at":"2026-07-05T10:12:08.116924Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:12:08.116924Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Ignore the KL Penalty! Boosting Exploration on Critical Tokens to Enhance RL Fine-Tuning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Jean Vassoyan, Nathana\\\"el Beau, Roman Plaud","submitted_at":"2025-02-10T14:56:25Z","abstract_excerpt":"The ability to achieve long-term goals is a key challenge in the current development of large language models (LLMs). To address this, pre-trained LLMs can be fine-tuned with reinforcement learning (RL) to explore solutions that optimize a given goal. However, exploration with LLMs is difficult, as a balance has to be struck between discovering new solutions and staying close enough to the pre-trained model, so as not to degrade basic capabilities. This is typically controlled with a Kullback-Leibler (KL) penalty. In this paper, we investigate the exploration dynamics of a small language model"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.06533","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/2502.06533/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":"2502.06533","created_at":"2026-07-05T10:12:08.116984+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.06533v1","created_at":"2026-07-05T10:12:08.116984+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.06533","created_at":"2026-07-05T10:12:08.116984+00:00"},{"alias_kind":"pith_short_12","alias_value":"W7J5BAKA7ZMB","created_at":"2026-07-05T10:12:08.116984+00:00"},{"alias_kind":"pith_short_16","alias_value":"W7J5BAKA7ZMBEENX","created_at":"2026-07-05T10:12:08.116984+00:00"},{"alias_kind":"pith_short_8","alias_value":"W7J5BAKA","created_at":"2026-07-05T10:12:08.116984+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2601.10348","citing_title":"Training-Trajectory-Aware Token Selection","ref_index":21,"is_internal_anchor":false},{"citing_arxiv_id":"2507.15778","citing_title":"Stabilizing Knowledge, Promoting Reasoning: Dual-Token Constraints for RLVR","ref_index":41,"is_internal_anchor":false},{"citing_arxiv_id":"2506.01939","citing_title":"Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08283","citing_title":"HTPO: Towards Exploration-Exploitation Balanced Policy Optimization via Hierarchical Token-level Objective Control","ref_index":34,"is_internal_anchor":false},{"citing_arxiv_id":"2604.16158","citing_title":"AtManRL: Towards Faithful Reasoning via Differentiable Attention Saliency","ref_index":17,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/W7J5BAKA7ZMBEENX4LM3KRP6IH","json":"https://pith.science/pith/W7J5BAKA7ZMBEENX4LM3KRP6IH.json","graph_json":"https://pith.science/api/pith-number/W7J5BAKA7ZMBEENX4LM3KRP6IH/graph.json","events_json":"https://pith.science/api/pith-number/W7J5BAKA7ZMBEENX4LM3KRP6IH/events.json","paper":"https://pith.science/paper/W7J5BAKA"},"agent_actions":{"view_html":"https://pith.science/pith/W7J5BAKA7ZMBEENX4LM3KRP6IH","download_json":"https://pith.science/pith/W7J5BAKA7ZMBEENX4LM3KRP6IH.json","view_paper":"https://pith.science/paper/W7J5BAKA","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.06533&json=true","fetch_graph":"https://pith.science/api/pith-number/W7J5BAKA7ZMBEENX4LM3KRP6IH/graph.json","fetch_events":"https://pith.science/api/pith-number/W7J5BAKA7ZMBEENX4LM3KRP6IH/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/W7J5BAKA7ZMBEENX4LM3KRP6IH/action/timestamp_anchor","attest_storage":"https://pith.science/pith/W7J5BAKA7ZMBEENX4LM3KRP6IH/action/storage_attestation","attest_author":"https://pith.science/pith/W7J5BAKA7ZMBEENX4LM3KRP6IH/action/author_attestation","sign_citation":"https://pith.science/pith/W7J5BAKA7ZMBEENX4LM3KRP6IH/action/citation_signature","submit_replication":"https://pith.science/pith/W7J5BAKA7ZMBEENX4LM3KRP6IH/action/replication_record"}},"created_at":"2026-07-05T10:12:08.116984+00:00","updated_at":"2026-07-05T10:12:08.116984+00:00"}