{"paper":{"title":"SaFeR-Steer: Evolving Multi-Turn MLLMs via Synthetic Bootstrapping and Feedback Dynamics","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"SaFeR-Steer trains multi-turn multimodal models to hold safety and helpfulness against escalating attacks through synthetic bootstrapping and tutor feedback.","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"An Zhang, Hanyu Li, Haolong Hu, Huahui Yi, Kun Wang, Qiankun Li, Tiancheng He, Yang Liu, Zhigang Zeng","submitted_at":"2026-03-18T09:28:29Z","abstract_excerpt":"MLLMs are increasingly deployed in multi-turn settings, where attackers can escalate unsafe intent through the evolving visual-text history and exploit long-context safety decay. Yet safety alignment is still dominated by single-turn data and fixed-template dialogues, leaving a mismatch between training and deployment.\n  To bridge this gap, we propose SaFeR-Steer, a progressive multi-turn alignment framework that combines staged synthetic bootstrapping with tutor-in-the-loop GRPO to train a single student under adaptive, on-policy attacks. We also introduce Trajectory-Consistent Summative Rewa"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Starting from Qwen2.5-VL-3B/7B, SaFeR-Steer substantially improves Safety/Helpfulness on both single-turn (48.30/45.86 -> 81.84/70.77 for 3B; 56.21/60.32 -> 87.89/77.40 for 7B) and multi-turn benchmarks (12.55/27.13 -> 55.58/70.27 for 3B; 24.66/46.48 -> 64.89/72.35 for 7B), shifting failures to later turns.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the synthetic bootstrapping process and tutor-in-the-loop GRPO produce data and gradients that generalize to real-world multi-turn attacks rather than overfitting to the generated distribution or the specific tutor model.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SaFeR-Steer boosts single-turn and multi-turn safety/helpfulness scores in Qwen2.5-VL models via staged synthetic bootstrapping and tutor-in-the-loop GRPO on a new STEER dataset.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"SaFeR-Steer trains multi-turn multimodal models to hold safety and helpfulness against escalating attacks through synthetic bootstrapping and tutor feedback.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"eeaf42adeb64867baf3def447b50c953ac971f612f45c14d63086572416a3fe8"},"source":{"id":"2604.16358","kind":"arxiv","version":2},"verdict":{"id":"e58ee4d9-d3c1-4f7d-be42-df15021160bf","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T09:46:57.358685Z","strongest_claim":"Starting from Qwen2.5-VL-3B/7B, SaFeR-Steer substantially improves Safety/Helpfulness on both single-turn (48.30/45.86 -> 81.84/70.77 for 3B; 56.21/60.32 -> 87.89/77.40 for 7B) and multi-turn benchmarks (12.55/27.13 -> 55.58/70.27 for 3B; 24.66/46.48 -> 64.89/72.35 for 7B), shifting failures to later turns.","one_line_summary":"SaFeR-Steer boosts single-turn and multi-turn safety/helpfulness scores in Qwen2.5-VL models via staged synthetic bootstrapping and tutor-in-the-loop GRPO on a new STEER dataset.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the synthetic bootstrapping process and tutor-in-the-loop GRPO produce data and gradients that generalize to real-world multi-turn attacks rather than overfitting to the generated distribution or the specific tutor model.","pith_extraction_headline":"SaFeR-Steer trains multi-turn multimodal models to hold safety and helpfulness against escalating attacks through synthetic bootstrapping and tutor feedback."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.16358/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"}