{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:SGTFOMT5U2N567NN2QIFNSYAUR","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"34b2f2138427fe72e67dea67a30b5e4808a9e4c70f93b18271adfbbecc62bd0b","cross_cats_sorted":["cs.AI"],"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.CL","submitted_at":"2026-02-17T14:46:48Z","title_canon_sha256":"b046897e425ba8f1ea06345a01ddc9783375117ae852d676dc9c4e1e7e9e70f1"},"schema_version":"1.0","source":{"id":"2602.15620","kind":"arxiv","version":5}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2602.15620","created_at":"2026-05-26T02:04:06Z"},{"alias_kind":"arxiv_version","alias_value":"2602.15620v5","created_at":"2026-05-26T02:04:06Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2602.15620","created_at":"2026-05-26T02:04:06Z"},{"alias_kind":"pith_short_12","alias_value":"SGTFOMT5U2N5","created_at":"2026-05-26T02:04:06Z"},{"alias_kind":"pith_short_16","alias_value":"SGTFOMT5U2N567NN","created_at":"2026-05-26T02:04:06Z"},{"alias_kind":"pith_short_8","alias_value":"SGTFOMT5","created_at":"2026-05-26T02:04:06Z"}],"graph_snapshots":[{"event_id":"sha256:e5318df5547c7f959f7df4710cf4106f0defa269c0d860dd1425b3eef30b7865","target":"graph","created_at":"2026-05-26T02:04:06Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Silencing gradients from a tiny fraction of spurious tokens stabilizes RL fine-tuning of LLMs and raises math reasoning performance."}],"snapshot_sha256":"5997830b62c025d9e4f9ea7945a5a9e00fdb91584d71bdb6f762a1928bef4e77"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"integrity":{"available":true,"clean":true,"detectors_run":[],"endpoint":"/pith/2602.15620/integrity.json","findings":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"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","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","cross_cats":["cs.AI"],"headline":"Silencing gradients from a tiny fraction of spurious tokens stabilizes RL fine-tuning of LLMs and raises math reasoning performance.","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.CL","submitted_at":"2026-02-17T14:46:48Z","title":"STAPO: Stabilizing Reinforcement Learning for LLMs by Silencing Rare Spurious Tokens"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2602.15620","kind":"arxiv","version":5},"verdict":{"created_at":"2026-05-15T21:38:10.962118Z","id":"28f91ef8-1e4e-447e-a0f7-066f663d1db5","model_set":{"reader":"grok-4.3"},"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","pith_extraction_headline":"Silencing gradients from a tiny fraction of spurious tokens stabilizes RL fine-tuning of LLMs and raises math reasoning performance.","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.","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."}},"verdict_id":"28f91ef8-1e4e-447e-a0f7-066f663d1db5"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:d485dfb1e5a50bd2ccb6c4e84eedb8ac6ab2b55585f731bf281311f6753c964f","target":"record","created_at":"2026-05-26T02:04:06Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"34b2f2138427fe72e67dea67a30b5e4808a9e4c70f93b18271adfbbecc62bd0b","cross_cats_sorted":["cs.AI"],"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.CL","submitted_at":"2026-02-17T14:46:48Z","title_canon_sha256":"b046897e425ba8f1ea06345a01ddc9783375117ae852d676dc9c4e1e7e9e70f1"},"schema_version":"1.0","source":{"id":"2602.15620","kind":"arxiv","version":5}},"canonical_sha256":"91a657327da69bdf7dadd41056cb00a44df0832c7bc88a360b85393174efae65","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"91a657327da69bdf7dadd41056cb00a44df0832c7bc88a360b85393174efae65","first_computed_at":"2026-05-26T02:04:06.525769Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-26T02:04:06.525769Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"TYrD7aZy9AvYdRNyunqUtG/2NHDDfiyR8s3xrS0Pmw57iBoLH5ula2ARHpnbDen5D+rGSA5Ya8UHyhMThQ7dBQ==","signature_status":"signed_v1","signed_at":"2026-05-26T02:04:06.526733Z","signed_message":"canonical_sha256_bytes"},"source_id":"2602.15620","source_kind":"arxiv","source_version":5}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:d485dfb1e5a50bd2ccb6c4e84eedb8ac6ab2b55585f731bf281311f6753c964f","sha256:e5318df5547c7f959f7df4710cf4106f0defa269c0d860dd1425b3eef30b7865"],"state_sha256":"95f9aee9f1a71a5ccbc8b49eecc4c292f6a3662e7f201da2bfda2064e66e5831"}