{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:RND6AARTW5VGVJBXVMQKBP7EAE","short_pith_number":"pith:RND6AART","schema_version":"1.0","canonical_sha256":"8b47e00233b76a6aa437ab20a0bfe40108d2cf81a3618cded7c76f9a75ad4785","source":{"kind":"arxiv","id":"2506.02177","version":1},"attestation_state":"computed","paper":{"title":"Act Only When It Pays: Efficient Reinforcement Learning for LLM Reasoning via Selective Rollouts","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.AI","authors_text":"Beidi Chen, Bhavya Kailkhura, Brian R. Bartoldson, Fan Lai, Haizhong Zheng, Jiawei Zhao, Yang Zhou","submitted_at":"2025-06-02T19:03:00Z","abstract_excerpt":"Reinforcement learning, such as PPO and GRPO, has powered recent breakthroughs in LLM reasoning. Scaling rollout to sample more prompts enables models to selectively use higher-quality data for training, which can stabilize RL training and improve model performance. However, this comes at the cost of significant computational overhead. In this paper, we show that a substantial portion of this overhead can be avoided by skipping uninformative prompts before rollout. Our analysis of reward dynamics reveals a strong temporal consistency in prompt value: prompts that are uninformative in one epoch"},"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":"2506.02177","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","primary_cat":"cs.AI","submitted_at":"2025-06-02T19:03:00Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"70002ab0940069bca99d1a2a4da6401eb1d0a531596b87ebff4d05de5c7a462c","abstract_canon_sha256":"59fa664d92ef60765eb31f1be3294fbc4c5ebb657074d459d3631020b70d083b"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:14:33.771488Z","signature_b64":"b6Ucmu8yj7IH5DwMMrrQdaO5TDVcjQym5eHntiwR9F0WAPqimWIGehvmq1iYkqi4OuQhaXektoxrkap3MjQrAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"8b47e00233b76a6aa437ab20a0bfe40108d2cf81a3618cded7c76f9a75ad4785","last_reissued_at":"2026-07-05T11:14:33.770986Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:14:33.770986Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Act Only When It Pays: Efficient Reinforcement Learning for LLM Reasoning via Selective Rollouts","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.AI","authors_text":"Beidi Chen, Bhavya Kailkhura, Brian R. Bartoldson, Fan Lai, Haizhong Zheng, Jiawei Zhao, Yang Zhou","submitted_at":"2025-06-02T19:03:00Z","abstract_excerpt":"Reinforcement learning, such as PPO and GRPO, has powered recent breakthroughs in LLM reasoning. Scaling rollout to sample more prompts enables models to selectively use higher-quality data for training, which can stabilize RL training and improve model performance. However, this comes at the cost of significant computational overhead. In this paper, we show that a substantial portion of this overhead can be avoided by skipping uninformative prompts before rollout. Our analysis of reward dynamics reveals a strong temporal consistency in prompt value: prompts that are uninformative in one epoch"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.02177","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/2506.02177/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":"2506.02177","created_at":"2026-07-05T11:14:33.771045+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.02177v1","created_at":"2026-07-05T11:14:33.771045+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.02177","created_at":"2026-07-05T11:14:33.771045+00:00"},{"alias_kind":"pith_short_12","alias_value":"RND6AARTW5VG","created_at":"2026-07-05T11:14:33.771045+00:00"},{"alias_kind":"pith_short_16","alias_value":"RND6AARTW5VGVJBX","created_at":"2026-07-05T11:14:33.771045+00:00"},{"alias_kind":"pith_short_8","alias_value":"RND6AART","created_at":"2026-07-05T11:14:33.771045+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":25,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.11634","citing_title":"Architecture-Aware Reinforcement Learning Makes Sliding-Window Attention Competitive in Math Reasoning","ref_index":51,"is_internal_anchor":false},{"citing_arxiv_id":"2606.10768","citing_title":"N-GRPO: Embedding-Level Neighbor Mixing for Enhanced Policy Optimization","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2606.08346","citing_title":"CATPO: Critique-Augmented Tree Policy Optimization","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06828","citing_title":"AdaGRPO: A Capability-Aware Adaptive Enhancement for Flow-based GRPO","ref_index":30,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06080","citing_title":"On Advantage Estimates for Max@K Policy Gradients","ref_index":69,"is_internal_anchor":false},{"citing_arxiv_id":"2606.01281","citing_title":"RLVR without Ineffective Samples: Group Prioritized Off-Policy Optimization for LLM Reasoning","ref_index":70,"is_internal_anchor":false},{"citing_arxiv_id":"2605.26606","citing_title":"Spend Your Rollouts Where It Counts: Rollout Allocation for Group-Based RL Post-Training","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00135","citing_title":"On Effectiveness and Efficiency of Agentic Tool-calling and RL Training","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06534","citing_title":"ROSE: Rollout On Serving GPUs via Cooperative Elasticity for Agentic RL","ref_index":94,"is_internal_anchor":false},{"citing_arxiv_id":"2605.15113","citing_title":"Learning from Language Feedback via Variational Policy Distillation","ref_index":51,"is_internal_anchor":false},{"citing_arxiv_id":"2605.15565","citing_title":"AstraFlow: Dataflow-Oriented Reinforcement Learning for Agentic LLMs","ref_index":47,"is_internal_anchor":false},{"citing_arxiv_id":"2509.02547","citing_title":"The Landscape of Agentic Reinforcement Learning for LLMs: A Survey","ref_index":57,"is_internal_anchor":false},{"citing_arxiv_id":"2605.13907","citing_title":"AIS: Adaptive Importance Sampling for Quantized RL","ref_index":26,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11235","citing_title":"Internalizing Curriculum Judgment for LLM Reinforcement Fine-Tuning","ref_index":16,"is_internal_anchor":false},{"citing_arxiv_id":"2604.28020","citing_title":"Cost-Aware Learning","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2605.08441","citing_title":"DUET: Optimize Token-Budget Allocation for Reinforcement Learning with Verifiable Rewards","ref_index":43,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07579","citing_title":"Your Language Model is Its Own Critic: Reinforcement Learning with Value Estimation from Actor's Internal States","ref_index":43,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09188","citing_title":"DARE: Difficulty-Adaptive Reinforcement Learning with Co-Evolved Difficulty Estimation","ref_index":61,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06534","citing_title":"ROSE: Rollout On Serving GPUs via Cooperative Elasticity for Agentic RL","ref_index":95,"is_internal_anchor":false},{"citing_arxiv_id":"2605.05802","citing_title":"Selective Rollout: Mid-Trajectory Termination for Multi-Sample Agent RL","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2604.06804","citing_title":"LASER: A Data-Centric Method for Low-Cost and Efficient SQL Rewriting based on SQL-GRPO","ref_index":49,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06755","citing_title":"Gradient Extrapolation-Based Policy Optimization","ref_index":45,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07579","citing_title":"Your Language Model is Its Own Critic: Reinforcement Learning with Value Estimation from Actor's Internal States","ref_index":43,"is_internal_anchor":false},{"citing_arxiv_id":"2605.02913","citing_title":"Generate, Filter, Control, Replay: A Comprehensive Survey of Rollout Strategies for LLM Reinforcement Learning","ref_index":184,"is_internal_anchor":false},{"citing_arxiv_id":"2604.16972","citing_title":"MCPO: Mastery-Consolidated Policy Optimization for Large Reasoning Models","ref_index":33,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/RND6AARTW5VGVJBXVMQKBP7EAE","json":"https://pith.science/pith/RND6AARTW5VGVJBXVMQKBP7EAE.json","graph_json":"https://pith.science/api/pith-number/RND6AARTW5VGVJBXVMQKBP7EAE/graph.json","events_json":"https://pith.science/api/pith-number/RND6AARTW5VGVJBXVMQKBP7EAE/events.json","paper":"https://pith.science/paper/RND6AART"},"agent_actions":{"view_html":"https://pith.science/pith/RND6AARTW5VGVJBXVMQKBP7EAE","download_json":"https://pith.science/pith/RND6AARTW5VGVJBXVMQKBP7EAE.json","view_paper":"https://pith.science/paper/RND6AART","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.02177&json=true","fetch_graph":"https://pith.science/api/pith-number/RND6AARTW5VGVJBXVMQKBP7EAE/graph.json","fetch_events":"https://pith.science/api/pith-number/RND6AARTW5VGVJBXVMQKBP7EAE/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/RND6AARTW5VGVJBXVMQKBP7EAE/action/timestamp_anchor","attest_storage":"https://pith.science/pith/RND6AARTW5VGVJBXVMQKBP7EAE/action/storage_attestation","attest_author":"https://pith.science/pith/RND6AARTW5VGVJBXVMQKBP7EAE/action/author_attestation","sign_citation":"https://pith.science/pith/RND6AARTW5VGVJBXVMQKBP7EAE/action/citation_signature","submit_replication":"https://pith.science/pith/RND6AARTW5VGVJBXVMQKBP7EAE/action/replication_record"}},"created_at":"2026-07-05T11:14:33.771045+00:00","updated_at":"2026-07-05T11:14:33.771045+00:00"}