{"_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_fb842e4846bf871ce1a802b154a015a28f3815421fe76aceef98ae232ef294be","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_fb842e4846bf871ce1a802b154a015a28f3815421fe76aceef98ae232ef294be","axiom_type":"AX.OBS","body":{"axiom_subtype":"news.vendor_press.v1","category":"news","fingerprint":"a1811fba2796a61596030981caa8cc82770ac38066544223d434e2acb5428cc4","published":"Tue, 30 Jun 2026 00:00:00 -0400","receipt_hash":"a1811fba2796a61596030981caa8cc82770ac38066544223d434e2acb5428cc4","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":"a1811fba2796a61596030981caa8cc82770ac38066544223d434e2acb5428cc4","observed_at":"2026-06-30T04:43:04.087680Z","parent_run_hash":"74f7ab392cc702044101fe24a76a2fdad11164cd79ce725aad6c446a477e89c5","published":"Tue, 30 Jun 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:2606.30627v1 Announce Type: cross \nAbstract: Conservative offline training is widely advocated as a safe foundation for subsequent online adaptation: if a policy stays close to well-supported behaviour, the argument goes, it is less likely to exploit imperfections in a learned reward model. We challenge this intuition empirically and mechanistically. We train a Qwen3-14B policy under Direct Preference Optimisation (DPO) with three levels of conservatism ($\\beta \\in \\{\\beta_{\\mathrm{lo}}, \\beta_{\\mathrm{mid}}, \\beta_{\\mathrm{hi}}\\}$ derived from empirical log-ratio percentiles), then adapt each checkpoint online against a learned reward ensemble (3\\,$\\times$\\,Qwen3-1.7B) while measuring true performance on GSM8K exact-answer accuracy. We find that \\emph{higher offline conservatism monotonically increases reward-hacking damage}, measured by the Goodhart gap and its area under the curve (AUGC), with Spearman $\\rho = 1.0$ across all three conditions. Mechanistic analysis reveals a th","title":"Pessimism's Paradox: Conservative Offline Training Amplifies Reward Hacking During Online Adaptation in Reasoning Models","url":"https://arxiv.org/abs/2606.30627","vendor":"arxiv_cs_ai"},"summary":"arXiv:2606.30627v1 Announce Type: cross \nAbstract: Conservative offline training is widely advocated as a safe foundation for subsequent online adaptation: if a policy stays close to well-supported behaviour, the argument goes, it is less likely to exploit imperfections in a learned reward model. We challenge this intuition empirically and mechanistically. We train a Qwen3-14B policy under Direct Preference Optimisation (DPO) with three levels of conservatism ($\\beta \\in \\{\\beta_{\\mathrm{lo}}, \\beta_{\\mathrm{mid}}, \\beta_{\\mathrm{hi}}\\}$ derived from empirical log-ratio percentiles), then adapt each checkpoint online against a learned reward ensemble (3\\,$\\times$\\,Qwen3-1.7B) while measuring true performance on GSM8K exact-answer accuracy. We find that \\emph{higher offline conservatism monotonically increases reward-hacking damage}, measured by the Goodhart gap and its area under the curve (AUGC), with Spearman $\\rho = 1.0$ across all three conditions. Mechanistic analysis reveals a th","title":"Pessimism's Paradox: Conservative Offline Training Amplifies Reward Hacking During Online Adaptation in Reasoning Models","vendor":"arxiv_cs_ai"},"confidence":{"method":"deterministic"},"decision":"POSITIVE","issued_at":"2026-06-30T04:43:04Z","notes":"Spider vendor_press (news) news.vendor_press.v1","object":{"captured_by":"crovia.spider.vendor_press","primary_source_url":"https://arxiv.org/abs/2606.30627"},"predecessors":[],"schema":"crovia.axiom.v1","signature":"ed25519:732f81a918180a1014f625ffc6b9947ffe2aebf7b4bfacbaf309ca3878e970c3a963295d618ed49b2c713619807cb2a1d1d63b35cb8aad5e1b62755b26132f09","signer":"crovia.substrate","subject":{"observed_at":"2026-06-30T04:43:04Z","source_collector":"spider:vendor_press","target_id":"https://arxiv.org/abs/2606.30627"},"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":"58ba3051016797527ea92b992e58a3918ecbb3391cdcb18128490123680b4c49","leaf_index":265034,"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":"08526ede68cf6af38b2de02022649d696e7db347886f254537a4dd73a2f2fc67","side":"right"},{"sibling":"138f60b836147657d17718ca4e8e4c46bf918c5e8b8645419b7d765fc056dd02","side":"left"},{"sibling":"712050da73506a2378045d99a893d701e5f90b702ac95fdab3ecb2a317c77c5b","side":"right"},{"sibling":"56ddfc750828425540ea79bedcf80245980ced6e42921e44bf734c5170076301","side":"left"},{"sibling":"80b2c038f2b521f28290fbce9fde21d0413e79f42a830c25d82739cb1e2e8610","side":"right"},{"sibling":"dc3a8a976dd0fb0b2dd04e514aa7cd1d06945d1036beb7fd4d0026c9c22682cb","side":"right"},{"sibling":"5b8a0910e4e94e2a9b10497e13f92efba841a089742fa095b3d09b024957b4bc","side":"left"},{"sibling":"e3ac74800103452465231dd60eaf9671a5039238ce21e35760b324bd56f70d4e","side":"right"},{"sibling":"3f337e26f1a0c4c64b8b7ebac04427325a7e7838c802c162576de1cd617b91ce","side":"left"},{"sibling":"1549a8883ab3267f958dc2624919e40c65c82958b8005967ed6a4da1247da0ad","side":"left"},{"sibling":"23994bf0974e5c9c7f63a61b4f0a48b0ca756a4adc34a8f85f878e774c37dfbe","side":"right"},{"sibling":"f9b4bed84fa6990c71ad2887c91bda183001f05f6b648f21d1273045a26b11fd","side":"left"},{"sibling":"173d2dc4b29ee04ea41d6d0ebc334c4bc2d46e7ee4230c94765413f24fb4bc42","side":"right"},{"sibling":"112461f7c0ec411116fb5c6c90fe95cea9d8f188b9fe08afe25a837ac02d0071","side":"right"},{"sibling":"ea9488204352c49db8f7daf05eefcd7628ecf9413830346674801a99d0654a94","side":"right"},{"sibling":"6261c13b9922cb657f10d1e5d36ec15d8771cf8766e36c61dcbffb7bed57e396","side":"right"},{"sibling":"fa19aa3faf287618b820bcfceebb366152ad521dd20ef9f51e977816663e448b","side":"right"},{"sibling":"c32f943406b62d1fc59b7f7e243492174c8e1caba8c8a2705f86c773315736e0","side":"right"},{"sibling":"1cecb7f447febd025aac272837c80de218aecc6485d2395a509b2a1f1b9c746e","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":265374,"merkle_root":"9636001ecab173cb6af10dc7c71eb14585daa62f9c0a6f027046f05633156891","public_key_hex":"cf742e26f75669dc673cb5c0786a1ae23ae8ca19c347317192ce40c28a7ff25c","run_id":"hourly_json_retrofit_20260630T053701Z","schema":"crovia.seal.v1","seal_family_version":"crovia-seal-family/1","seal_kind":"substrate_batch","sealed_at":"2026-06-30T05:38:03Z","sig_algorithm":"ed25519","signature":"6d4b8fd9b9da856cbb5fba7540877c6a63fa18a5ec3eaf28dc4d6d1c64921c9c0565f8c95f4b7aec0e7744fd7754844f051baf863db708cad768765b416a7b0c","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_fb842e4846bf871ce1a802b154a015a28f3815421fe76aceef98ae232ef294be"}}