{"_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_45395495ff7bb287da6cce89b8e9443535f55e22b27fdc96146040be4ccbdeaf","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_45395495ff7bb287da6cce89b8e9443535f55e22b27fdc96146040be4ccbdeaf","axiom_type":"AX.OBS","body":{"axiom_subtype":"news.vendor_press.v1","category":"news","fingerprint":"d5518dc784ae605baebf0c790ea88d93aeca3b8735c3a48b6239b6c3128fc2ff","published":"Tue, 28 Jul 2026 00:00:00 -0400","receipt_hash":"d5518dc784ae605baebf0c790ea88d93aeca3b8735c3a48b6239b6c3128fc2ff","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":"d5518dc784ae605baebf0c790ea88d93aeca3b8735c3a48b6239b6c3128fc2ff","observed_at":"2026-07-28T04:43:08.282317Z","parent_run_hash":"23a1ef85134515049ced29518443d084afc46fd7c967741e6c6acdbdbbf29939","published":"Tue, 28 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:2607.23394v1 Announce Type: new \nAbstract: Recent work shows that fine-tuning language models on even a small amount of poisoned data can install targeted misbehavior, and ostensibly benign data can transmit hidden preferences that generalize broadly. Standard defenses, such as data filtering, mixing in harmless data, and regularization, attenuate these effects but do not eliminate them. We instead pursue robustness through redundancy: collecting multiple datasets from different sources and only learning what is common between them. Thus, if only a subset of sources are malicious, the misbehavior will be blocked. In order to implement this defense strategy, we fine-tune a separate reference model on each source's dataset and aggregate their next-token distributions at decoding time. We introduce two consensus decoders: a token-wise minimum, which caps each token at the lowest probability any source assigns, and a base-relative variant, which reverts to the base probability on any","title":"Inference-Time Consensus for Mitigating Hidden Behaviors from LLM Fine-Tuning","url":"https://arxiv.org/abs/2607.23394","vendor":"arxiv_cs_ai"},"summary":"arXiv:2607.23394v1 Announce Type: new \nAbstract: Recent work shows that fine-tuning language models on even a small amount of poisoned data can install targeted misbehavior, and ostensibly benign data can transmit hidden preferences that generalize broadly. Standard defenses, such as data filtering, mixing in harmless data, and regularization, attenuate these effects but do not eliminate them. We instead pursue robustness through redundancy: collecting multiple datasets from different sources and only learning what is common between them. Thus, if only a subset of sources are malicious, the misbehavior will be blocked. In order to implement this defense strategy, we fine-tune a separate reference model on each source's dataset and aggregate their next-token distributions at decoding time. We introduce two consensus decoders: a token-wise minimum, which caps each token at the lowest probability any source assigns, and a base-relative variant, which reverts to the base probability on any","title":"Inference-Time Consensus for Mitigating Hidden Behaviors from LLM Fine-Tuning","vendor":"arxiv_cs_ai"},"confidence":{"method":"deterministic"},"decision":"POSITIVE","issued_at":"2026-07-28T04:43:08Z","notes":"Spider vendor_press (news) news.vendor_press.v1","object":{"captured_by":"crovia.spider.vendor_press","primary_source_url":"https://arxiv.org/abs/2607.23394"},"predecessors":[],"schema":"crovia.axiom.v1","signature":"ed25519:75db41c667fbcc1eecaa3344d1ee7ee31caff292994f6b7389885e833ab6c00fc2c826682b629a1867e7ea6101d3ccbec17b7514cd1b2a3f3b74081d3d926b0a","signer":"crovia.substrate","subject":{"observed_at":"2026-07-28T04:43:08Z","source_collector":"spider:vendor_press","target_id":"https://arxiv.org/abs/2607.23394"},"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":"d47df189d747dccf069871cd037f51613e0217c4a3817a659fcf78c00149e494","leaf_index":360424,"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":"7ecedcbef48760375d627ee7c742a39bfe2b2945d610b46c25643625d48a399f","side":"right"},{"sibling":"087d076edaf9bb79208992cc0e7f0097567ee437607267f09048fe23dd171942","side":"right"},{"sibling":"715d4f5a13829919def51657cfb56e2c7bd171ceb73cd9dcb5b234bb1f1a6e92","side":"right"},{"sibling":"d91bf25868d267899a659e74b97b19b9b530c31c1af525a2e95ef7129b2e538f","side":"left"},{"sibling":"f5c3fc065d7cabfcb497904e608f1ce9972da8ce33157b38e8c8a18ef457b596","side":"right"},{"sibling":"4c5cb02f387999a1cf3a76135da2a320a4bff24236d08d026d50804da6ccf567","side":"left"},{"sibling":"d686aadab346134cea3b1a9ca407741e39c2faaec0ce19ee820e7526a4ebd400","side":"left"},{"sibling":"ce393210d8cb9e0705a24ff80bd76d7becf22bfce1258f4a57fd2dbdca17218b","side":"left"},{"sibling":"8fe957ea378915d410087f8b2c41171bdfdac25ba34c42d93c5aaa835f32246c","side":"left"},{"sibling":"f57ebea8749a327b4323e9ad9c906a9400f5a5d8c83084dfc7d80be83bd44007","side":"left"},{"sibling":"4424ff61f20c9cef2251672611ec86369d87b1608ab089738930fd3d52e0dd54","side":"left"},{"sibling":"8db22b6d8b4004df8ccd40f648d9fd8a62f564821c80fbfd9c743849f4102e1b","side":"left"},{"sibling":"cff500acb83b14a8a7195a72d90dd4b7a6b8a28fc1069f8300b1191728718f90","side":"left"},{"sibling":"51ee2566d84aba54b9d07233fb460f9fab044b4edfa3fbe376fa1df0727acfa5","side":"left"},{"sibling":"f3e45bceed774d2402fa45d41ff5190f295823bd2f216eb90157884150034693","side":"left"},{"sibling":"2096cd69b54e283ddc45b26b63564a5e4c02f303ffd855b3b3533bcf26bc284d","side":"right"},{"sibling":"77025bcb374a7ad74f520643e20a8ae1205a7ee78507b0beb93117760f1c29d3","side":"left"},{"sibling":"3b50864499c874394ea0928567747666eaf59b01380e46cd52164ec5acec0f71","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":361008,"merkle_root":"3065e8369ea437c06beba806dc4e4bb159979adeb21fe632242c1906a7204647","public_key_hex":"cf742e26f75669dc673cb5c0786a1ae23ae8ca19c347317192ce40c28a7ff25c","run_id":"hourly_json_retrofit_20260728T053701Z","schema":"crovia.seal.v1","seal_family_version":"crovia-seal-family/1","seal_kind":"substrate_batch","sealed_at":"2026-07-28T05:38:48Z","sig_algorithm":"ed25519","signature":"9141644407577a82611c1579110f667de2d46dc6b93c6322edf26f4c3056ea99f0e56502853908e30d87c38bcf95eb6e0ab5130525aa51505bd6f61938120609","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_45395495ff7bb287da6cce89b8e9443535f55e22b27fdc96146040be4ccbdeaf"}}