{"paper":{"title":"A Model Can Help Itself: Reward-Free Self-Training for LLM Reasoning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Language models can improve their reasoning by training on responses they generate themselves.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Anthony Man-Cho So, Lei Zhao, Mengqi Li, Ruoyu Sun, Xiao Li","submitted_at":"2025-10-21T17:15:56Z","abstract_excerpt":"Can language models improve their reasoning performance without external rewards, using only their own sampled responses for training? We show that they can. We propose Self-evolving Post-Training (SePT), a simple post-training method that alternates between self-generation and training on self-generated responses. It repeatedly samples questions, uses the model itself to generate responses under a specified sampling temperature, and then trains the model on the self-generated data. In this self-training loop, we use an online data refresh mechanism, where each new batch is generated by the mo"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We show that they can. We propose Self-evolving Post-Training (SePT), a simple post-training method that alternates between self-generation and training on self-generated responses. Across six math reasoning benchmarks, SePT improves a strong no-training baseline... and in some settings can even approach the performance of Reinforcement Learning with Verifiable Rewards (RLVR).","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That self-generated responses supply a net positive training signal rather than reinforcing the model's existing errors or hallucinations, which must hold for the iterative self-training loop to produce sustained gains without external verification or filtering.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SePT enables LLMs to improve math reasoning on multiple benchmarks by iteratively training on their own low-temperature generated responses using an online data refresh mechanism.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Language models can improve their reasoning by training on responses they generate themselves.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"d1eefac338f9d4b3286071d0d1691c85c584bb514eff26205aa22a72cc14b1da"},"source":{"id":"2510.18814","kind":"arxiv","version":3},"verdict":{"id":"f4d25a43-0b64-42de-9fb7-be4d16c84530","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-18T05:07:29.377760Z","strongest_claim":"We show that they can. We propose Self-evolving Post-Training (SePT), a simple post-training method that alternates between self-generation and training on self-generated responses. Across six math reasoning benchmarks, SePT improves a strong no-training baseline... and in some settings can even approach the performance of Reinforcement Learning with Verifiable Rewards (RLVR).","one_line_summary":"SePT enables LLMs to improve math reasoning on multiple benchmarks by iteratively training on their own low-temperature generated responses using an online data refresh mechanism.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That self-generated responses supply a net positive training signal rather than reinforcing the model's existing errors or hallucinations, which must hold for the iterative self-training loop to produce sustained gains without external verification or filtering.","pith_extraction_headline":"Language models can improve their reasoning by training on responses they generate themselves."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2510.18814/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":2,"snapshot_sha256":"1da77da5bf2e75e62ba903468a8c56f50941e63395d2864ab68fa7a509218b65"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}