{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:SK2RJZGVJJOO2FS5TEBCKP4JGF","short_pith_number":"pith:SK2RJZGV","schema_version":"1.0","canonical_sha256":"92b514e4d54a5ced165d9902253f89314679cc1bff91d10eb9c2370efecc2ac2","source":{"kind":"arxiv","id":"2603.25184","version":2},"attestation_state":"computed","paper":{"title":"Train at Moving Edge: Online-Verified Prompt Selection for Efficient RL Training of Large Reasoning Model","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Bailong Lin, Chen Jason Zhang, Jiahao Wu, Ke Tang, Kun Wang, Li Qing, Ning Lu, Shengcai Liu, Yanting Yang","submitted_at":"2026-03-26T08:52:35Z","abstract_excerpt":"Reinforcement learning (RL) has become essential for post-training large language models (LLMs) in reasoning tasks. While scaling rollouts can stabilize training and enhance performance, the computational overhead is a critical issue. In algorithms like GRPO, multiple rollouts per prompt incur prohibitive costs, as a large portion of prompts provide negligible gradients and are thus of low utility. To address this problem, we investigate how to select high-utility prompts before the rollout phase. Our experimental analysis reveals that sample utility is non-uniform and evolving: the strongest "},"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":"2603.25184","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-26T08:52:35Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"7d58a8e04616bab81f1eaaa340378803edfa163d6099dec9167fd307361d0417","abstract_canon_sha256":"e47ce6d9aad31d393ed0bb71f3c5422e5ed79465865358ec4782ac09bfeb6e57"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-09T01:05:16.525699Z","signature_b64":"1AmgFp2W9J8FVo85GiHF723db0FmAlRZY36LForZS9H3TPJGlqqWqb1+pKhFXPpYm48+392DPyoTMwy/V8/8CQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"92b514e4d54a5ced165d9902253f89314679cc1bff91d10eb9c2370efecc2ac2","last_reissued_at":"2026-06-09T01:05:16.525251Z","signature_status":"signed_v1","first_computed_at":"2026-06-09T01:05:16.525251Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Train at Moving Edge: Online-Verified Prompt Selection for Efficient RL Training of Large Reasoning Model","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Bailong Lin, Chen Jason Zhang, Jiahao Wu, Ke Tang, Kun Wang, Li Qing, Ning Lu, Shengcai Liu, Yanting Yang","submitted_at":"2026-03-26T08:52:35Z","abstract_excerpt":"Reinforcement learning (RL) has become essential for post-training large language models (LLMs) in reasoning tasks. While scaling rollouts can stabilize training and enhance performance, the computational overhead is a critical issue. In algorithms like GRPO, multiple rollouts per prompt incur prohibitive costs, as a large portion of prompts provide negligible gradients and are thus of low utility. To address this problem, we investigate how to select high-utility prompts before the rollout phase. Our experimental analysis reveals that sample utility is non-uniform and evolving: the strongest "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2603.25184","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/2603.25184/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":"2603.25184","created_at":"2026-06-09T01:05:16.525326+00:00"},{"alias_kind":"arxiv_version","alias_value":"2603.25184v2","created_at":"2026-06-09T01:05:16.525326+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2603.25184","created_at":"2026-06-09T01:05:16.525326+00:00"},{"alias_kind":"pith_short_12","alias_value":"SK2RJZGVJJOO","created_at":"2026-06-09T01:05:16.525326+00:00"},{"alias_kind":"pith_short_16","alias_value":"SK2RJZGVJJOO2FS5","created_at":"2026-06-09T01:05:16.525326+00:00"},{"alias_kind":"pith_short_8","alias_value":"SK2RJZGV","created_at":"2026-06-09T01:05:16.525326+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2605.08756","citing_title":"AHD Agent: Agentic Reinforcement Learning for Automatic Heuristic Design","ref_index":43,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/SK2RJZGVJJOO2FS5TEBCKP4JGF","json":"https://pith.science/pith/SK2RJZGVJJOO2FS5TEBCKP4JGF.json","graph_json":"https://pith.science/api/pith-number/SK2RJZGVJJOO2FS5TEBCKP4JGF/graph.json","events_json":"https://pith.science/api/pith-number/SK2RJZGVJJOO2FS5TEBCKP4JGF/events.json","paper":"https://pith.science/paper/SK2RJZGV"},"agent_actions":{"view_html":"https://pith.science/pith/SK2RJZGVJJOO2FS5TEBCKP4JGF","download_json":"https://pith.science/pith/SK2RJZGVJJOO2FS5TEBCKP4JGF.json","view_paper":"https://pith.science/paper/SK2RJZGV","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2603.25184&json=true","fetch_graph":"https://pith.science/api/pith-number/SK2RJZGVJJOO2FS5TEBCKP4JGF/graph.json","fetch_events":"https://pith.science/api/pith-number/SK2RJZGVJJOO2FS5TEBCKP4JGF/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/SK2RJZGVJJOO2FS5TEBCKP4JGF/action/timestamp_anchor","attest_storage":"https://pith.science/pith/SK2RJZGVJJOO2FS5TEBCKP4JGF/action/storage_attestation","attest_author":"https://pith.science/pith/SK2RJZGVJJOO2FS5TEBCKP4JGF/action/author_attestation","sign_citation":"https://pith.science/pith/SK2RJZGVJJOO2FS5TEBCKP4JGF/action/citation_signature","submit_replication":"https://pith.science/pith/SK2RJZGVJJOO2FS5TEBCKP4JGF/action/replication_record"}},"created_at":"2026-06-09T01:05:16.525326+00:00","updated_at":"2026-06-09T01:05:16.525326+00:00"}