{"_canonicalization":{"envelope_id":"axm_ + sha256(envelope minus {signature, axiom_id, anchors})","envelope_signature":"ed25519(envelope minus {signature, axiom_id})","json":"sort_keys=True, separators=(',',':'), ensure_ascii=False, allow_nan=False, utf-8","leaf_hash":"sha256(0x00 || canonical_json(envelope_full))","seal_signature":"ed25519(seal minus {signature, sig_algorithm})"},"axiom_id":"axm_235b54b32661156ae792a927b1981c643892b5bd3ebf1fb26864cc13ec4dd4b5","bitcoin_anchor":{"bitcoin_attestations":[],"calendar_attestations":[],"ots_url":"","stamped_at":"","status":"pending_next_stamp"},"envelope":{"anchors":[{"chain":"crovia.axiom_graph","height":0,"merkle_proof":"spider_vendor_press_v1","root_at_anchor":"spider_vendor_press_v1"}],"axiom_id":"axm_235b54b32661156ae792a927b1981c643892b5bd3ebf1fb26864cc13ec4dd4b5","axiom_type":"AX.OBS","body":{"axiom_subtype":"news.vendor_press.v1","category":"news","fingerprint":"3ef595b8512f419c6806e1e2f156982136001a2c5bfd032287a6f47d6b450918","published":"Mon, 18 May 2026 00:00:00 -0400","receipt_hash":"3ef595b8512f419c6806e1e2f156982136001a2c5bfd032287a6f47d6b450918","schema":"spider.news.vendor_press.v1","spider":"vendor_press","spider_record":{"axiom_subtype":"news.vendor_press.v1","category":"news","decision_hint":"POSITIVE","envelope_target":"AX.OBS","fingerprint":"3ef595b8512f419c6806e1e2f156982136001a2c5bfd032287a6f47d6b450918","observed_at":"2026-05-18T04:43:11.219741Z","parent_run_hash":"a8aad7414ebb6b75c726f09cd673410576a7f87e191fbdb9ddac99e9b2b95a05","published":"Mon, 18 May 2026 00:00:00 -0400","runtime_version":"0.1.0","schema":"spider.news.vendor_press.v1","source_status":200,"source_url":"https://export.arxiv.org/rss/cs.AI","spider":"vendor_press","summary_excerpt":"arXiv:2605.12483v3 Announce Type: replace-cross \nAbstract: We present a four-stage post-training workflow for LLM reasoning that allocates scarce labeled training data more effectively than standard recipes. The stages are: (1) sparse-reward RL on a larger teacher; (2a) forward-KL warmup on teacher rollouts; (2b) on-policy distillation under student rollouts; (3) optional sparse-reward RL on the deployment student using any held-out labeled data. On verifiable math with a Qwen3-1.7B deployment student, the workflow reaches $79.3\\%$ MATH and $25.2\\%$ AIME~2024 (avg@16), versus $75.9\\%$ and $19.8\\%$ for direct GRPO on the same student. We justify the workflow through a reward-density principle: each gradient step of on-policy distillation is a local trust-region update under a dense teacher-induced implicit reward, informative only when the teacher is itself reward-shaped (condition C1) and lies within a small KL of the student (condition C2). Stages~1 and~2a are the explicit devices tha","title":"Beyond GRPO and On-Policy Distillation: An Empirical Sparse-to-Dense Reward Principle for Language-Model Post-Training","url":"https://arxiv.org/abs/2605.12483","vendor":"arxiv_cs_ai"},"summary":"arXiv:2605.12483v3 Announce Type: replace-cross \nAbstract: We present a four-stage post-training workflow for LLM reasoning that allocates scarce labeled training data more effectively than standard recipes. The stages are: (1) sparse-reward RL on a larger teacher; (2a) forward-KL warmup on teacher rollouts; (2b) on-policy distillation under student rollouts; (3) optional sparse-reward RL on the deployment student using any held-out labeled data. On verifiable math with a Qwen3-1.7B deployment student, the workflow reaches $79.3\\%$ MATH and $25.2\\%$ AIME~2024 (avg@16), versus $75.9\\%$ and $19.8\\%$ for direct GRPO on the same student. We justify the workflow through a reward-density principle: each gradient step of on-policy distillation is a local trust-region update under a dense teacher-induced implicit reward, informative only when the teacher is itself reward-shaped (condition C1) and lies within a small KL of the student (condition C2). Stages~1 and~2a are the explicit devices tha","title":"Beyond GRPO and On-Policy Distillation: An Empirical Sparse-to-Dense Reward Principle for Language-Model Post-Training","vendor":"arxiv_cs_ai"},"confidence":{"method":"deterministic"},"decision":"POSITIVE","issued_at":"2026-05-18T04:43:11Z","notes":"Spider vendor_press (news) news.vendor_press.v1","object":{"captured_by":"crovia.spider.vendor_press","primary_source_url":"https://arxiv.org/abs/2605.12483"},"predecessors":[],"schema":"crovia.axiom.v1","signature":"ed25519:ebd8751400bcef34ac62ee60c9be35e02c1598819df91ffd8357360fd34282b85fc837402d3d6bc65ec5e2d1cb33171c37c5d39d88822749f7aa68e0eb3bf50f","signer":"crovia.substrate","subject":{"observed_at":"2026-05-18T04:43:11Z","source_collector":"spider:vendor_press","target_id":"https://arxiv.org/abs/2605.12483"},"tsa":{"authority":"crovia.substrate.bootstrap","rfc3161_token":"{\"kind\":\"crovia.bootstrap.tsa\",\"source_jsonl\":\"/opt/crovia/spider/data/news/vendor_press_v1.jsonl\",\"source_seal_merkle_root\":\"spider_vendor_press_v1\",\"upgrade_path\":\"Sessione H \\u2014 OpenTimestamps weekly anchor\"}"},"zk_mode":"clear","zk_proof":null},"ledger":{"leaf_hash":"a6c8af2320d906d9fb2c674ec9d34b5924a7f47a97a2a7a9bba06ea4cb794a9b","leaf_index":140827,"ledger_path":"/opt/crovia/substrate/axiom_ledger.jsonl"},"merkle_proof":{"hash_alg":"sha256","leaf_prefix":"0x00","node_prefix":"0x01","odd_leaf_rule":"duplicate_last","path":[{"sibling":"040d0df11d88bec49d672470e971cc5bee6f53936af7bcb4143cf2c602b5e7f4","side":"left"},{"sibling":"2388a1c0d7712c215b7f006d485d14d1a51fc1a0e2475eb1a5695fd82d1029a7","side":"left"},{"sibling":"e1a4cdfdb833946468e77a5a81f060a8cd99bf095368dbaf5d79c17c881a49a0","side":"right"},{"sibling":"d6d61752f75ee3ab22ea04deb01e5850f268f534e86b179c86eebec7c2254453","side":"left"},{"sibling":"f727513692aa02658ee9539f81f537f0867b56b99c709dd467221cf4364e3efa","side":"left"},{"sibling":"2c1eb16147433fe66739128d4f460e8c2a75eee663b7d9dc9c3ad70cfc8c0429","side":"right"},{"sibling":"f5549d405a3735c35bfcb3acd74cae69f4d03370836eec8bbc18e7b3eb016a19","side":"right"},{"sibling":"a87432691ceb26cb93d306a629d23c791a2630efd7fc708d3ad480b3ef31e966","side":"right"},{"sibling":"eec7831167f92e73ff0d6defbc9a917c62b82a9f14cdfbb761db0e7219ab3519","side":"right"},{"sibling":"796add816cf5ab3ce5f338195252cc064685523572fa7fad3982a6e20730395f","side":"left"},{"sibling":"28b78fb112bcf26b6801664db97eb8f52a9bccbf0a7ae6766e11845d443692df","side":"left"},{"sibling":"68d0a4634c1460a19c92edd9480df3aa733b814463e7420d1e14471bf61b2f83","side":"right"},{"sibling":"8af64f275b862349aa3bbb9d5cd7fa9a7fdd5620af3bf1b36b2a4519b0b53bdf","side":"right"},{"sibling":"8f4c0fbe56b6c010fbb8c782ebcd478079bb3f991d8704e2534209a075d9163c","side":"left"},{"sibling":"b98c2afadb358e5387e88f19588f8343a81b488d9b44a6f7e57a032db3a1b030","side":"right"},{"sibling":"11b0c1591747f09f7c8971a6caa19befcd81317ca9dfd417b143234df4e10c79","side":"right"},{"sibling":"87206f3bcc342797c990d87f7235c01f78d32ca59cfaf8ad18d71afc879ba477","side":"right"},{"sibling":"d841ad93efda0869e5eb97678f348f03f5caab4353e05ff4bf18f47fb945b822","side":"left"}]},"schema":"crovia.axiom_proof.v1","seal":{"first_collector_run_id":"","first_receipt_hash":"","jsonl_path":"/opt/crovia/substrate/axiom_ledger.jsonl","key_id":"430895f101d38164","last_collector_run_id":"","last_receipt_hash":"","leaf_count":140892,"merkle_root":"6cca56ead155990456b8a014cc50bddbe710f409b26e3d1bfa6fb12b0bfcf6bf","public_key_hex":"cf742e26f75669dc673cb5c0786a1ae23ae8ca19c347317192ce40c28a7ff25c","run_id":"hourly_json_retrofit_20260518T053701Z","schema":"crovia.seal.v1","seal_family_version":"crovia-seal-family/1","seal_kind":"substrate_batch","sealed_at":"2026-05-18T05:37:30Z","sig_algorithm":"ed25519","signature":"1e1135f7595f79b14fb11f5fa81a2e17ad31b11b44b427a5e40a7d511cd86447daf492babd368ab571cf26404c8c74c450d460130fca4b064eb2760367489a0f","signer_version":"1.1.0"},"trust_root":{"key_id":"430895f101d38164","public_key_hex":"cf742e26f75669dc673cb5c0786a1ae23ae8ca19c347317192ce40c28a7ff25c","signature_algorithm":"ed25519","url":"/registry/canon/TRUST_ROOT.md"},"verifier":{"spec":"/registry/canon/AXIOM_RECEIPT_v1.md","url":"/v/axm_235b54b32661156ae792a927b1981c643892b5bd3ebf1fb26864cc13ec4dd4b5"}}