{"paper":{"title":"STAPO: Stabilizing Reinforcement Learning for LLMs by Silencing Rare Spurious Tokens","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"Silencing gradients from a tiny fraction of spurious tokens stabilizes RL fine-tuning of LLMs and raises math reasoning performance.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Bo Zhang, Guojian Zhan, Jiang Wu, Jingliang Duan, Kehua Sheng, Keqiang Li, Letian Tao, Shengbo Eben Li, Shiqi Liu, Yang Guan, Yinuo Wang, Zeyu He, Zhilong Zheng","submitted_at":"2026-02-17T14:46:48Z","abstract_excerpt":"Reinforcement Learning (RL) has significantly improved large language model reasoning, but existing RL fine-tuning methods rely heavily on heuristic techniques such as entropy regularization and reweighting to maintain stability. In practice, they often suffer from late-stage performance collapse, leading to degraded reasoning quality and unstable training. We identify a key factor behind this instability: a small fraction of tokens, termed spurious tokens (around 0.01%), which contribute little to the reasoning outcome but receive disproportionately amplified gradient updates due to inheritin"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across six mathematical reasoning benchmarks using Qwen 1.7B, 8B, and 14B base models, STAPO consistently demonstrates superior entropy stability and achieves an average performance improvement of 11.49% (ρ_T=1.0, top-p=1.0) and 3.73% (ρ_T=0.7, top-p=0.9) over GRPO, 20-Entropy, and JustRL.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the identified spurious tokens (0.01% fraction) are the primary driver of instability and that silencing their gradients does not discard useful reasoning signal or introduce new biases in the policy update.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"STAPO stabilizes RL for LLMs by suppressing gradient updates from rare spurious tokens, yielding 11.49% average gains on math benchmarks over GRPO and similar baselines.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Silencing gradients from a tiny fraction of spurious tokens stabilizes RL fine-tuning of LLMs and raises math reasoning performance.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"5997830b62c025d9e4f9ea7945a5a9e00fdb91584d71bdb6f762a1928bef4e77"},"source":{"id":"2602.15620","kind":"arxiv","version":5},"verdict":{"id":"28f91ef8-1e4e-447e-a0f7-066f663d1db5","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T21:38:10.962118Z","strongest_claim":"Across six mathematical reasoning benchmarks using Qwen 1.7B, 8B, and 14B base models, STAPO consistently demonstrates superior entropy stability and achieves an average performance improvement of 11.49% (ρ_T=1.0, top-p=1.0) and 3.73% (ρ_T=0.7, top-p=0.9) over GRPO, 20-Entropy, and JustRL.","one_line_summary":"STAPO stabilizes RL for LLMs by suppressing gradient updates from rare spurious tokens, yielding 11.49% average gains on math benchmarks over GRPO and similar baselines.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the identified spurious tokens (0.01% fraction) are the primary driver of instability and that silencing their gradients does not discard useful reasoning signal or introduce new biases in the policy update.","pith_extraction_headline":"Silencing gradients from a tiny fraction of spurious tokens stabilizes RL fine-tuning of LLMs and raises math reasoning performance."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2602.15620/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"}