{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:NBSTXQN4XVAB7LB7AL6TSKM7AU","short_pith_number":"pith:NBSTXQN4","schema_version":"1.0","canonical_sha256":"68653bc1bcbd401fac3f02fd39299f0526c60be074053abb8825498f40c774c8","source":{"kind":"arxiv","id":"2507.17107","version":2},"attestation_state":"computed","paper":{"title":"Reinforcement Learning Fine-Tunes a Sparse Subnetwork in Large Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Andrii Balashov","submitted_at":"2025-07-23T01:02:17Z","abstract_excerpt":"Reinforcement learning (RL) is a key post-pretraining step for aligning large language models (LLMs) with complex tasks and human preferences. While it is often assumed that RL fine-tuning requires updating most of a model's parameters, we challenge this assumption with a surprising finding: RL fine-tuning consistently modifies only a small subnetwork (typically 5-30% of weights), leaving most parameters unchanged. We call this phenomenon RL-induced parameter update sparsity. It arises naturally, without any sparsity constraints or parameter-efficient tuning, and appears across multiple RL alg"},"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":"2507.17107","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-07-23T01:02:17Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"d80c336febb51f6e4d0229c28d38c9cb3d32cb52f65d9faf7c4d78e659d157fe","abstract_canon_sha256":"02de09453dd20d24c85b32f63c9f20a328a07a208c94e8129c90148e01f83bae"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:44:52.183250Z","signature_b64":"AHEMNM1HDI7P7TxIB8R425kxDEixYOpf63lNwytPs4cWZQWzuBE9NcpdO/EvhVQKWZomHmIeR0BKO2KdL6JoAw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"68653bc1bcbd401fac3f02fd39299f0526c60be074053abb8825498f40c774c8","last_reissued_at":"2026-07-05T11:44:52.182768Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:44:52.182768Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Reinforcement Learning Fine-Tunes a Sparse Subnetwork in Large Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Andrii Balashov","submitted_at":"2025-07-23T01:02:17Z","abstract_excerpt":"Reinforcement learning (RL) is a key post-pretraining step for aligning large language models (LLMs) with complex tasks and human preferences. While it is often assumed that RL fine-tuning requires updating most of a model's parameters, we challenge this assumption with a surprising finding: RL fine-tuning consistently modifies only a small subnetwork (typically 5-30% of weights), leaving most parameters unchanged. We call this phenomenon RL-induced parameter update sparsity. It arises naturally, without any sparsity constraints or parameter-efficient tuning, and appears across multiple RL alg"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2507.17107","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/2507.17107/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":"2507.17107","created_at":"2026-07-05T11:44:52.182836+00:00"},{"alias_kind":"arxiv_version","alias_value":"2507.17107v2","created_at":"2026-07-05T11:44:52.182836+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2507.17107","created_at":"2026-07-05T11:44:52.182836+00:00"},{"alias_kind":"pith_short_12","alias_value":"NBSTXQN4XVAB","created_at":"2026-07-05T11:44:52.182836+00:00"},{"alias_kind":"pith_short_16","alias_value":"NBSTXQN4XVAB7LB7","created_at":"2026-07-05T11:44:52.182836+00:00"},{"alias_kind":"pith_short_8","alias_value":"NBSTXQN4","created_at":"2026-07-05T11:44:52.182836+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/NBSTXQN4XVAB7LB7AL6TSKM7AU","json":"https://pith.science/pith/NBSTXQN4XVAB7LB7AL6TSKM7AU.json","graph_json":"https://pith.science/api/pith-number/NBSTXQN4XVAB7LB7AL6TSKM7AU/graph.json","events_json":"https://pith.science/api/pith-number/NBSTXQN4XVAB7LB7AL6TSKM7AU/events.json","paper":"https://pith.science/paper/NBSTXQN4"},"agent_actions":{"view_html":"https://pith.science/pith/NBSTXQN4XVAB7LB7AL6TSKM7AU","download_json":"https://pith.science/pith/NBSTXQN4XVAB7LB7AL6TSKM7AU.json","view_paper":"https://pith.science/paper/NBSTXQN4","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2507.17107&json=true","fetch_graph":"https://pith.science/api/pith-number/NBSTXQN4XVAB7LB7AL6TSKM7AU/graph.json","fetch_events":"https://pith.science/api/pith-number/NBSTXQN4XVAB7LB7AL6TSKM7AU/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/NBSTXQN4XVAB7LB7AL6TSKM7AU/action/timestamp_anchor","attest_storage":"https://pith.science/pith/NBSTXQN4XVAB7LB7AL6TSKM7AU/action/storage_attestation","attest_author":"https://pith.science/pith/NBSTXQN4XVAB7LB7AL6TSKM7AU/action/author_attestation","sign_citation":"https://pith.science/pith/NBSTXQN4XVAB7LB7AL6TSKM7AU/action/citation_signature","submit_replication":"https://pith.science/pith/NBSTXQN4XVAB7LB7AL6TSKM7AU/action/replication_record"}},"created_at":"2026-07-05T11:44:52.182836+00:00","updated_at":"2026-07-05T11:44:52.182836+00:00"}