{"_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_5c5b4c366ca9aa07c64860c81479c0c688625f8f24f0d94a252c578c95c1163a","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_5c5b4c366ca9aa07c64860c81479c0c688625f8f24f0d94a252c578c95c1163a","axiom_type":"AX.OBS","body":{"axiom_subtype":"news.vendor_press.v1","category":"news","fingerprint":"52faaf94c769c7e0a55ae6a8b83d2ab66f078e67ad46fa35c02d7b29a3c1a90e","published":"Mon, 13 Jul 2026 00:00:00 -0400","receipt_hash":"52faaf94c769c7e0a55ae6a8b83d2ab66f078e67ad46fa35c02d7b29a3c1a90e","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":"52faaf94c769c7e0a55ae6a8b83d2ab66f078e67ad46fa35c02d7b29a3c1a90e","observed_at":"2026-07-13T04:43:08.394955Z","parent_run_hash":"900c1c934245e788564e199a9619f2dc36ec91d9804ddd9c6a40fb42c8a1e1c0","published":"Mon, 13 Jul 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:2607.09492v1 Announce Type: new \nAbstract: Reinforcement learning (RL) is increasingly used to align multimodal large language models (MLLMs), but higher rewards do not always imply better task performance. This risk is amplified when visual evidence is evaluated by text-only or weakly grounded rewards. We study reward hacking in MLLM RL across safety VQA, chart VQA, and stress-test settings, varying reward design, data ambiguity, model scale (2B-32B), and RL algorithm (GRPO, RLOO, DAPO). We introduce Newly Rewarded Failure Rate (NRFR), which measures failures among samples whose proxy reward improves over the SFT baseline. Outcome-only rewards cause severe hacking, reaching 48.1% Reward Hacking Rate (RHR), while NRFR exceeding RHR shows that RL creates new failures rather than merely inheriting them. Scaling reduces but does not eliminate hacking: even the 32B model retains a 54.9% worse rate under outcome-only rewards, whereas answer-aware rewards improve the oracle trend at ev","title":"Multimodal Reward Hacking in Reinforcement Learning","url":"https://arxiv.org/abs/2607.09492","vendor":"arxiv_cs_ai"},"summary":"arXiv:2607.09492v1 Announce Type: new \nAbstract: Reinforcement learning (RL) is increasingly used to align multimodal large language models (MLLMs), but higher rewards do not always imply better task performance. This risk is amplified when visual evidence is evaluated by text-only or weakly grounded rewards. We study reward hacking in MLLM RL across safety VQA, chart VQA, and stress-test settings, varying reward design, data ambiguity, model scale (2B-32B), and RL algorithm (GRPO, RLOO, DAPO). We introduce Newly Rewarded Failure Rate (NRFR), which measures failures among samples whose proxy reward improves over the SFT baseline. Outcome-only rewards cause severe hacking, reaching 48.1% Reward Hacking Rate (RHR), while NRFR exceeding RHR shows that RL creates new failures rather than merely inheriting them. Scaling reduces but does not eliminate hacking: even the 32B model retains a 54.9% worse rate under outcome-only rewards, whereas answer-aware rewards improve the oracle trend at ev","title":"Multimodal Reward Hacking in Reinforcement Learning","vendor":"arxiv_cs_ai"},"confidence":{"method":"deterministic"},"decision":"POSITIVE","issued_at":"2026-07-13T04:43:08Z","notes":"Spider vendor_press (news) news.vendor_press.v1","object":{"captured_by":"crovia.spider.vendor_press","primary_source_url":"https://arxiv.org/abs/2607.09492"},"predecessors":[],"schema":"crovia.axiom.v1","signature":"ed25519:946045649fb4c0f885e0c8c64f6cf2c70eab35b0a257ac67ac92ba2968b3315d793ca0d1212195b8a8743c3e7481b3335f4f772482ff493ac3e7f22fa0d49a01","signer":"crovia.substrate","subject":{"observed_at":"2026-07-13T04:43:08Z","source_collector":"spider:vendor_press","target_id":"https://arxiv.org/abs/2607.09492"},"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":"7921abca217f5f1f7a9b637e437768c463707094bcad6d80a4cccaf82395ffab","leaf_index":309589,"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":"ea654a7e8fa1a4be0ec17dcbf0d09beeb47daf827c29955df5804eda45b9106a","side":"left"},{"sibling":"83905a81ec3fa9617a02bd8203d9a1b1a022eefddd5ad14cafe9e271c59e2afe","side":"right"},{"sibling":"b8617f9b109502540d6e257328f5f3e6b482848bfed73a5d31b06c392fcffb32","side":"left"},{"sibling":"2d2c277c63752d74fe691f57339bda6dc20a8c25648560951018e019dfb35806","side":"right"},{"sibling":"d3b622e4dce65dfa8a21d298383c33df9d707c81ff7748a178576b771302bb7b","side":"left"},{"sibling":"44441d791f015edb1b6a6a65e55acd1a326e31793948509b417f1d8fe38b87fe","side":"right"},{"sibling":"2b24156380d32b87505f9ff8abbd321b7fb6de2890cac94ceb9223f26203bff0","side":"left"},{"sibling":"1ffbbb6083898ec6beabcbd39595cc5a65c28e8ba2a8441a35ec731530072d34","side":"right"},{"sibling":"6cbb0c4695e74fa8ec17dfde89fa56c1958a1d4dffbbcbe3556da4a69c699ebc","side":"left"},{"sibling":"c4bde3283be97905033d39c1097ac73b82f180de8830384fe6f9f1933f7fa4b9","side":"right"},{"sibling":"3b5f968ebea87e7be458ef7e63c6637988d27ba6d170e1f3794d385cd76ca23f","side":"right"},{"sibling":"5ed534e945b31085c140b50460415e7960b76a4b6266da67d272b853dc94b352","side":"left"},{"sibling":"91010b271bc8eb5253b3549292ef3146e1d85bc9bac7ca36e0862f1204f84e8d","side":"left"},{"sibling":"772fbc112e94d8e574379343387c65503d3cf8fb16ff89090174eccced871a41","side":"left"},{"sibling":"5d50450cae1a230f682b390c8e28ae822ec6c0c04a9bc79b0d27af97306ccccd","side":"right"},{"sibling":"d8b9143917b539c543cf4448cec00131f8b807bd8004979c54ebe09798748c66","side":"left"},{"sibling":"e9ac4b1e71d3f572751b7984e9ca00893d71628137b4634e82df02cf3db9ab68","side":"right"},{"sibling":"99ff86058e045249bf936a629308be31e4cc71328f0282c4a924f4e6718be5f0","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":309862,"merkle_root":"f18a76abb66e7cb448986b6541091416ed4ebdde8124f5208a7c7c94bb4165d1","public_key_hex":"cf742e26f75669dc673cb5c0786a1ae23ae8ca19c347317192ce40c28a7ff25c","run_id":"hourly_json_retrofit_20260713T053701Z","schema":"crovia.seal.v1","seal_family_version":"crovia-seal-family/1","seal_kind":"substrate_batch","sealed_at":"2026-07-13T05:38:23Z","sig_algorithm":"ed25519","signature":"0488e3527b1d559ba55116220f91e2f6c22bb5358a135e8a9b94a65450b50511702044fa993555662ebca09f005b277f5f6ad0d044ec96a5568e9993c09ef007","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_5c5b4c366ca9aa07c64860c81479c0c688625f8f24f0d94a252c578c95c1163a"}}