{"paper":{"title":"Model Spec Midtraining: Improving How Alignment Training Generalizes","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Training models on synthetic documents about their Model Spec before alignment fine-tuning shapes how they generalize from later examples.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Chloe Li, Jon Kutasov, Nevan Wichers, Samuel Marks, Sara Price","submitted_at":"2026-05-03T23:16:14Z","abstract_excerpt":"Some frontier AI developers aim to align language models to a Model Spec or Constitution that describes the intended model behavior. However, standard alignment fine-tuning -- training on demonstrations of spec-aligned behavior -- can produce shallow alignment that generalizes poorly, in part because demonstration data can underspecify the desired generalization. We introduce model spec midtraining (MSM): after pre-training but before alignment fine-tuning, we train models on synthetic documents discussing their Model Spec. This teaches models the content of the spec, thereby shaping how they "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"applying MSM with a spec addressing self-preservation and goal-guarding substantially reduces agentic misalignment rate (Qwen3-32B: 54% to 7%), beating a deliberative alignment baseline (14%).","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That training on synthetic documents discussing the Model Spec will reliably encode the intended generalizations into the model without introducing new unintended behaviors or degrading other capabilities.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Model spec midtraining trains AI models on documents about their alignment rules before demonstration fine-tuning, producing stronger and more controllable generalization to the intended values and safety behaviors.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Training models on synthetic documents about their Model Spec before alignment fine-tuning shapes how they generalize from later examples.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"7f628945b51e438d630ee4d9120457419a2a2519e440f7c4a3725ecadeb3277d"},"source":{"id":"2605.02087","kind":"arxiv","version":2},"verdict":{"id":"dc57bf1f-cdbd-45c8-bb9f-eb5e43c3efda","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-08T19:10:59.393102Z","strongest_claim":"applying MSM with a spec addressing self-preservation and goal-guarding substantially reduces agentic misalignment rate (Qwen3-32B: 54% to 7%), beating a deliberative alignment baseline (14%).","one_line_summary":"Model spec midtraining trains AI models on documents about their alignment rules before demonstration fine-tuning, producing stronger and more controllable generalization to the intended values and safety behaviors.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That training on synthetic documents discussing the Model Spec will reliably encode the intended generalizations into the model without introducing new unintended behaviors or degrading other capabilities.","pith_extraction_headline":"Training models on synthetic documents about their Model Spec before alignment fine-tuning shapes how they generalize from later examples."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.02087/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T16:38:00.882396Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-20T04:01:22.783057Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T16:40:46.733574Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"c9036558d6cb50e7cd8492ddc1f6f4422dc4bfc4f2c3f3b9aa044e511877a68d"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"cd9373de7701a1add7448ff08ff170df1bab33080f297e6571a9ffa56834bff4"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}