{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:PEDOHQUVII2HU5SNFXCIXLGTOK","short_pith_number":"pith:PEDOHQUV","schema_version":"1.0","canonical_sha256":"7906e3c29542347a764d2dc48bacd3729be76ac3039d9399da5726f6a41b6208","source":{"kind":"arxiv","id":"2504.11343","version":2},"attestation_state":"computed","paper":{"title":"A Minimalist Approach to LLM Reasoning: from Rejection Sampling to Reinforce","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL","stat.ML"],"primary_cat":"cs.LG","authors_text":"Bo Pang, Caiming Xiong, Doyen Sahoo, HanZe Dong, Jiarui Yao, Junnan Li, Lei Wang, Nan Jiang, Tong Zhang, Wei Xiong, Yuhui Xu","submitted_at":"2025-04-15T16:15:02Z","abstract_excerpt":"Reinforcement learning (RL) has become a prevailing approach for fine-tuning large language models (LLMs) on complex reasoning tasks. Among recent methods, GRPO stands out for its empirical success in training models such as DeepSeek-R1, yet the sources of its effectiveness remain poorly understood. In this work, we revisit GRPO from a reinforce-like algorithm perspective and analyze its core components. Surprisingly, we find that a simple rejection sampling baseline, RAFT, which trains only on positively rewarded samples, yields competitive performance than GRPO and PPO. Our ablation studies "},"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":"2504.11343","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-04-15T16:15:02Z","cross_cats_sorted":["cs.AI","cs.CL","stat.ML"],"title_canon_sha256":"b9614bf8c6c1274359ef5d71e92d0659f929406ea8ec25e8942f69d6cbdaf043","abstract_canon_sha256":"12856dad416e61d51a39dd018080245ba9ebe1346b5018e39f7d3ed5ec5f1109"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:20:11.363503Z","signature_b64":"xMi4E7da4aR+ekmUgfAb2mmxyj1KmF2Rs2gQZq6XqosNcBTHniymYLK8zctC/WspGn/OuMx/JliMBNi7oNCYBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"7906e3c29542347a764d2dc48bacd3729be76ac3039d9399da5726f6a41b6208","last_reissued_at":"2026-07-05T11:20:11.362227Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:20:11.362227Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"A Minimalist Approach to LLM Reasoning: from Rejection Sampling to Reinforce","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL","stat.ML"],"primary_cat":"cs.LG","authors_text":"Bo Pang, Caiming Xiong, Doyen Sahoo, HanZe Dong, Jiarui Yao, Junnan Li, Lei Wang, Nan Jiang, Tong Zhang, Wei Xiong, Yuhui Xu","submitted_at":"2025-04-15T16:15:02Z","abstract_excerpt":"Reinforcement learning (RL) has become a prevailing approach for fine-tuning large language models (LLMs) on complex reasoning tasks. Among recent methods, GRPO stands out for its empirical success in training models such as DeepSeek-R1, yet the sources of its effectiveness remain poorly understood. In this work, we revisit GRPO from a reinforce-like algorithm perspective and analyze its core components. Surprisingly, we find that a simple rejection sampling baseline, RAFT, which trains only on positively rewarded samples, yields competitive performance than GRPO and PPO. Our ablation studies "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2504.11343","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/2504.11343/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":"2504.11343","created_at":"2026-07-05T11:20:11.362899+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.11343v2","created_at":"2026-07-05T11:20:11.362899+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.11343","created_at":"2026-07-05T11:20:11.362899+00:00"},{"alias_kind":"pith_short_12","alias_value":"PEDOHQUVII2H","created_at":"2026-07-05T11:20:11.362899+00:00"},{"alias_kind":"pith_short_16","alias_value":"PEDOHQUVII2HU5SN","created_at":"2026-07-05T11:20:11.362899+00:00"},{"alias_kind":"pith_short_8","alias_value":"PEDOHQUV","created_at":"2026-07-05T11:20:11.362899+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":26,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.07646","citing_title":"RL Post-Training Builds Compositional Reasoning Strategies","ref_index":11,"is_internal_anchor":true},{"citing_arxiv_id":"2606.22570","citing_title":"What are Key Factors for Updates in RL for LLM Reasoning?","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2606.18089","citing_title":"From Reasoning Traces to Reusable Modules: Understanding Compositional Generalization in Language Model Reasoning","ref_index":137,"is_internal_anchor":false},{"citing_arxiv_id":"2606.18307","citing_title":"DRIFT: Refining Instruction Data via On-Policy Data Attribution","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2606.05597","citing_title":"AsyncWebRL: Efficient Asynchronous Reinforcement Learning for Multi-Step Visual Web Agents","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2606.04560","citing_title":"Rollout-Level Advantage-Prioritized Experience Replay for GRPO","ref_index":67,"is_internal_anchor":false},{"citing_arxiv_id":"2606.02812","citing_title":"Traj-Evolve: A Self-Evolving Multi-Agent System for Patient Trajectory Modeling in Lung Cancer Early Detection","ref_index":46,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00257","citing_title":"ARCA: Adapter-Residual Credit Assignment When Token Signals Degenerate","ref_index":26,"is_internal_anchor":false},{"citing_arxiv_id":"2504.13818","citing_title":"Not All Rollouts are Useful: Down-Sampling Rollouts in LLM Reinforcement Learning","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2505.17086","citing_title":"Advancing Multi-Agent RAG Systems with Minimalist Reinforcement Learning","ref_index":86,"is_internal_anchor":false},{"citing_arxiv_id":"2509.03403","citing_title":"Beyond Correctness: Harmonizing Process and Outcome Rewards through RL Training","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2605.20865","citing_title":"Multi-Step Likelihood-Ratio Correction for Reinforcement Learning with Verifiable Rewards","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2604.28005","citing_title":"Kernelized Advantage Estimation: From Nonparametric Statistics to LLM Reasoning","ref_index":26,"is_internal_anchor":false},{"citing_arxiv_id":"2508.20697","citing_title":"Token Buncher: Shielding LLMs from Harmful Reinforcement Learning Fine-Tuning","ref_index":59,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11505","citing_title":"Selective Off-Policy Reference Tuning with Plan Guidance","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11726","citing_title":"Block-R1: Rethinking the Role of Block Size in Multi-domain Reinforcement Learning for Diffusion Large Language Models","ref_index":65,"is_internal_anchor":false},{"citing_arxiv_id":"2509.16117","citing_title":"DiffusionNFT: Online Diffusion Reinforcement with Forward Process","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11726","citing_title":"Block-R1: Rethinking the Role of Block Size in Multi-domain Reinforcement Learning for Diffusion Large Language Models","ref_index":65,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11505","citing_title":"Selective Off-Policy Reference Tuning with Plan Guidance","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2604.28005","citing_title":"Kernelized Advantage Estimation: From Nonparametric Statistics to LLM Reasoning","ref_index":26,"is_internal_anchor":false},{"citing_arxiv_id":"2604.26326","citing_title":"Addressing Performance Saturation for LLM RL via Precise Entropy Curve Control","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2604.26326","citing_title":"Addressing Performance Saturation for LLM RL via Precise Entropy Curve Control","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2605.05965","citing_title":"Beyond Uniform Credit Assignment: Selective Eligibility Traces for RLVR","ref_index":54,"is_internal_anchor":false},{"citing_arxiv_id":"2605.00347","citing_title":"Odysseus: Scaling VLMs to 100+ Turn Decision-Making in Games via Reinforcement Learning","ref_index":77,"is_internal_anchor":false},{"citing_arxiv_id":"2605.07244","citing_title":"Experience Sharing in Mutual Reinforcement Learning for Heterogeneous Language Models","ref_index":99,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/PEDOHQUVII2HU5SNFXCIXLGTOK","json":"https://pith.science/pith/PEDOHQUVII2HU5SNFXCIXLGTOK.json","graph_json":"https://pith.science/api/pith-number/PEDOHQUVII2HU5SNFXCIXLGTOK/graph.json","events_json":"https://pith.science/api/pith-number/PEDOHQUVII2HU5SNFXCIXLGTOK/events.json","paper":"https://pith.science/paper/PEDOHQUV"},"agent_actions":{"view_html":"https://pith.science/pith/PEDOHQUVII2HU5SNFXCIXLGTOK","download_json":"https://pith.science/pith/PEDOHQUVII2HU5SNFXCIXLGTOK.json","view_paper":"https://pith.science/paper/PEDOHQUV","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.11343&json=true","fetch_graph":"https://pith.science/api/pith-number/PEDOHQUVII2HU5SNFXCIXLGTOK/graph.json","fetch_events":"https://pith.science/api/pith-number/PEDOHQUVII2HU5SNFXCIXLGTOK/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/PEDOHQUVII2HU5SNFXCIXLGTOK/action/timestamp_anchor","attest_storage":"https://pith.science/pith/PEDOHQUVII2HU5SNFXCIXLGTOK/action/storage_attestation","attest_author":"https://pith.science/pith/PEDOHQUVII2HU5SNFXCIXLGTOK/action/author_attestation","sign_citation":"https://pith.science/pith/PEDOHQUVII2HU5SNFXCIXLGTOK/action/citation_signature","submit_replication":"https://pith.science/pith/PEDOHQUVII2HU5SNFXCIXLGTOK/action/replication_record"}},"created_at":"2026-07-05T11:20:11.362899+00:00","updated_at":"2026-07-05T11:20:11.362899+00:00"}