{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:3DAI3M6PTNBQY6FQJUWNEV6MWI","short_pith_number":"pith:3DAI3M6P","schema_version":"1.0","canonical_sha256":"d8c08db3cf9b430c78b04d2cd257ccb22024449f130e7593e1abac08e3823214","source":{"kind":"arxiv","id":"2605.28467","version":1},"attestation_state":"computed","paper":{"title":"Mitigating Adaptive Attacks against Reasoning Models with Activation Consistency Training","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Avidan Shah, Jannik Brinkmann, Rico Angell","submitted_at":"2026-05-27T13:33:26Z","abstract_excerpt":"As LLMs gain stronger reasoning capabilities, their extended chain-of-thought introduces new degrees of complexity for defending against adversarial jailbreaks and prompt injection. We study consistency training, a family of fine-tuning objectives that enforce identical behavior on clean prompts and adversarial rewrites, and evaluate its two main variants, output-level (BCT) and activation-level (ACT), across five reasoning models. We formulate both methods as a prompt injection defense and find ACT to be competitive with other training-based defenses while requiring only self-supervised pairs"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2605.28467","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-05-27T13:33:26Z","cross_cats_sorted":[],"title_canon_sha256":"a47abab0963673886cf60bb7ce5e6059c18dbc77e47eebc6110d6ab32a389497","abstract_canon_sha256":"e8871745350361d057329eda5a4f05e34e8ec93e3cfba4aafe0a9f07b00eea22"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-28T02:04:53.880815Z","signature_b64":"dtR8ScuGh2JBEShRHvGAvWwfQrI12FaVckxfv3LsL8budlchJ+iiyKS41rdutwKCQq48ZJ6VNZ+uxhc9thPQAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d8c08db3cf9b430c78b04d2cd257ccb22024449f130e7593e1abac08e3823214","last_reissued_at":"2026-05-28T02:04:53.880429Z","signature_status":"signed_v1","first_computed_at":"2026-05-28T02:04:53.880429Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Mitigating Adaptive Attacks against Reasoning Models with Activation Consistency Training","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Avidan Shah, Jannik Brinkmann, Rico Angell","submitted_at":"2026-05-27T13:33:26Z","abstract_excerpt":"As LLMs gain stronger reasoning capabilities, their extended chain-of-thought introduces new degrees of complexity for defending against adversarial jailbreaks and prompt injection. We study consistency training, a family of fine-tuning objectives that enforce identical behavior on clean prompts and adversarial rewrites, and evaluate its two main variants, output-level (BCT) and activation-level (ACT), across five reasoning models. We formulate both methods as a prompt injection defense and find ACT to be competitive with other training-based defenses while requiring only self-supervised pairs"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.28467","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.28467/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2605.28467","created_at":"2026-05-28T02:04:53.880484+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.28467v1","created_at":"2026-05-28T02:04:53.880484+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.28467","created_at":"2026-05-28T02:04:53.880484+00:00"},{"alias_kind":"pith_short_12","alias_value":"3DAI3M6PTNBQ","created_at":"2026-05-28T02:04:53.880484+00:00"},{"alias_kind":"pith_short_16","alias_value":"3DAI3M6PTNBQY6FQ","created_at":"2026-05-28T02:04:53.880484+00:00"},{"alias_kind":"pith_short_8","alias_value":"3DAI3M6P","created_at":"2026-05-28T02:04:53.880484+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/3DAI3M6PTNBQY6FQJUWNEV6MWI","json":"https://pith.science/pith/3DAI3M6PTNBQY6FQJUWNEV6MWI.json","graph_json":"https://pith.science/api/pith-number/3DAI3M6PTNBQY6FQJUWNEV6MWI/graph.json","events_json":"https://pith.science/api/pith-number/3DAI3M6PTNBQY6FQJUWNEV6MWI/events.json","paper":"https://pith.science/paper/3DAI3M6P"},"agent_actions":{"view_html":"https://pith.science/pith/3DAI3M6PTNBQY6FQJUWNEV6MWI","download_json":"https://pith.science/pith/3DAI3M6PTNBQY6FQJUWNEV6MWI.json","view_paper":"https://pith.science/paper/3DAI3M6P","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.28467&json=true","fetch_graph":"https://pith.science/api/pith-number/3DAI3M6PTNBQY6FQJUWNEV6MWI/graph.json","fetch_events":"https://pith.science/api/pith-number/3DAI3M6PTNBQY6FQJUWNEV6MWI/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3DAI3M6PTNBQY6FQJUWNEV6MWI/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3DAI3M6PTNBQY6FQJUWNEV6MWI/action/storage_attestation","attest_author":"https://pith.science/pith/3DAI3M6PTNBQY6FQJUWNEV6MWI/action/author_attestation","sign_citation":"https://pith.science/pith/3DAI3M6PTNBQY6FQJUWNEV6MWI/action/citation_signature","submit_replication":"https://pith.science/pith/3DAI3M6PTNBQY6FQJUWNEV6MWI/action/replication_record"}},"created_at":"2026-05-28T02:04:53.880484+00:00","updated_at":"2026-05-28T02:04:53.880484+00:00"}