{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:PSKIOVTH3OSE3VZCPTMMBIBFOW","short_pith_number":"pith:PSKIOVTH","canonical_record":{"source":{"id":"2604.16358","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-18T09:28:29Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"72f46ab6d5f941f24e43ae73fcf421145ce251a8b0b3013c6b065c1aabaaeefa","abstract_canon_sha256":"01084c11191a11914693f030ba9811d9905f01d853b535a544793ac371bc6d3c"},"schema_version":"1.0"},"canonical_sha256":"7c94875667dba44dd7227cd8c0a0257580243f89bf2e37c6043c5ed0a04e4bde","source":{"kind":"arxiv","id":"2604.16358","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2604.16358","created_at":"2026-05-28T02:04:47Z"},{"alias_kind":"arxiv_version","alias_value":"2604.16358v2","created_at":"2026-05-28T02:04:47Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2604.16358","created_at":"2026-05-28T02:04:47Z"},{"alias_kind":"pith_short_12","alias_value":"PSKIOVTH3OSE","created_at":"2026-05-28T02:04:47Z"},{"alias_kind":"pith_short_16","alias_value":"PSKIOVTH3OSE3VZC","created_at":"2026-05-28T02:04:47Z"},{"alias_kind":"pith_short_8","alias_value":"PSKIOVTH","created_at":"2026-05-28T02:04:47Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:PSKIOVTH3OSE3VZCPTMMBIBFOW","target":"record","payload":{"canonical_record":{"source":{"id":"2604.16358","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-18T09:28:29Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"72f46ab6d5f941f24e43ae73fcf421145ce251a8b0b3013c6b065c1aabaaeefa","abstract_canon_sha256":"01084c11191a11914693f030ba9811d9905f01d853b535a544793ac371bc6d3c"},"schema_version":"1.0"},"canonical_sha256":"7c94875667dba44dd7227cd8c0a0257580243f89bf2e37c6043c5ed0a04e4bde","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-28T02:04:47.904341Z","signature_b64":"nCUjOkuGECr3j0M1bMIse5cYOjNBx3geVGkTQ9paZqKPc6s82qZL9eM0jJhrdZmdRrt2OONVuDdMhf2Te80eBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"7c94875667dba44dd7227cd8c0a0257580243f89bf2e37c6043c5ed0a04e4bde","last_reissued_at":"2026-05-28T02:04:47.903904Z","signature_status":"signed_v1","first_computed_at":"2026-05-28T02:04:47.903904Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2604.16358","source_version":2,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-28T02:04:47Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"958V6cFh9p7vYXjxPs7PNJgkyM33zlQuSnuvNrH+NLfudAwElQCZ7hySliLaFhKmDk7MWlOkwkeUT5g8CwKkBQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-08T16:46:06.076910Z"},"content_sha256":"19cf3f55b2b1f545b8c3163f7db1c847dd0218f0d3d1ab33c56debefbb9071aa","schema_version":"1.0","event_id":"sha256:19cf3f55b2b1f545b8c3163f7db1c847dd0218f0d3d1ab33c56debefbb9071aa"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:PSKIOVTH3OSE3VZCPTMMBIBFOW","target":"graph","payload":{"graph_snapshot":{"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"},"verdict_id":"e58ee4d9-d3c1-4f7d-be42-df15021160bf"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-28T02:04:47Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"9sPoND3TmXsk9oneAgI51qkK1en3C0Y6OWLZgLYfJEavHWbHe7Daa2dSThygByLiQ/Z0VniML1ykIbpV+K6+Dg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-08T16:46:06.077894Z"},"content_sha256":"c8158fd9e84cc92d6edc49fbe9d5b323580bb100649e6efdc00ca8b27730083d","schema_version":"1.0","event_id":"sha256:c8158fd9e84cc92d6edc49fbe9d5b323580bb100649e6efdc00ca8b27730083d"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/PSKIOVTH3OSE3VZCPTMMBIBFOW/bundle.json","state_url":"https://pith.science/pith/PSKIOVTH3OSE3VZCPTMMBIBFOW/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/PSKIOVTH3OSE3VZCPTMMBIBFOW/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-06-08T16:46:06Z","links":{"resolver":"https://pith.science/pith/PSKIOVTH3OSE3VZCPTMMBIBFOW","bundle":"https://pith.science/pith/PSKIOVTH3OSE3VZCPTMMBIBFOW/bundle.json","state":"https://pith.science/pith/PSKIOVTH3OSE3VZCPTMMBIBFOW/state.json","well_known_bundle":"https://pith.science/.well-known/pith/PSKIOVTH3OSE3VZCPTMMBIBFOW/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:PSKIOVTH3OSE3VZCPTMMBIBFOW","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"01084c11191a11914693f030ba9811d9905f01d853b535a544793ac371bc6d3c","cross_cats_sorted":["cs.CL"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-18T09:28:29Z","title_canon_sha256":"72f46ab6d5f941f24e43ae73fcf421145ce251a8b0b3013c6b065c1aabaaeefa"},"schema_version":"1.0","source":{"id":"2604.16358","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2604.16358","created_at":"2026-05-28T02:04:47Z"},{"alias_kind":"arxiv_version","alias_value":"2604.16358v2","created_at":"2026-05-28T02:04:47Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2604.16358","created_at":"2026-05-28T02:04:47Z"},{"alias_kind":"pith_short_12","alias_value":"PSKIOVTH3OSE","created_at":"2026-05-28T02:04:47Z"},{"alias_kind":"pith_short_16","alias_value":"PSKIOVTH3OSE3VZC","created_at":"2026-05-28T02:04:47Z"},{"alias_kind":"pith_short_8","alias_value":"PSKIOVTH","created_at":"2026-05-28T02:04:47Z"}],"graph_snapshots":[{"event_id":"sha256:c8158fd9e84cc92d6edc49fbe9d5b323580bb100649e6efdc00ca8b27730083d","target":"graph","created_at":"2026-05-28T02:04:47Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","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."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"SaFeR-Steer trains multi-turn multimodal models to hold safety and helpfulness against escalating attacks through synthetic bootstrapping and tutor feedback."}],"snapshot_sha256":"eeaf42adeb64867baf3def447b50c953ac971f612f45c14d63086572416a3fe8"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"integrity":{"available":true,"clean":true,"detectors_run":[],"endpoint":"/pith/2604.16358/integrity.json","findings":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"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","authors_text":"An Zhang, Hanyu Li, Haolong Hu, Huahui Yi, Kun Wang, Qiankun Li, Tiancheng He, Yang Liu, Zhigang Zeng","cross_cats":["cs.CL"],"headline":"SaFeR-Steer trains multi-turn multimodal models to hold safety and helpfulness against escalating attacks through synthetic bootstrapping and tutor feedback.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-18T09:28:29Z","title":"SaFeR-Steer: Evolving Multi-Turn MLLMs via Synthetic Bootstrapping and Feedback Dynamics"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2604.16358","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-15T09:46:57.358685Z","id":"e58ee4d9-d3c1-4f7d-be42-df15021160bf","model_set":{"reader":"grok-4.3"},"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","pith_extraction_headline":"SaFeR-Steer trains multi-turn multimodal models to hold safety and helpfulness against escalating attacks through synthetic bootstrapping and tutor feedback.","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.","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."}},"verdict_id":"e58ee4d9-d3c1-4f7d-be42-df15021160bf"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:19cf3f55b2b1f545b8c3163f7db1c847dd0218f0d3d1ab33c56debefbb9071aa","target":"record","created_at":"2026-05-28T02:04:47Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"01084c11191a11914693f030ba9811d9905f01d853b535a544793ac371bc6d3c","cross_cats_sorted":["cs.CL"],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2026-03-18T09:28:29Z","title_canon_sha256":"72f46ab6d5f941f24e43ae73fcf421145ce251a8b0b3013c6b065c1aabaaeefa"},"schema_version":"1.0","source":{"id":"2604.16358","kind":"arxiv","version":2}},"canonical_sha256":"7c94875667dba44dd7227cd8c0a0257580243f89bf2e37c6043c5ed0a04e4bde","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"7c94875667dba44dd7227cd8c0a0257580243f89bf2e37c6043c5ed0a04e4bde","first_computed_at":"2026-05-28T02:04:47.903904Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-28T02:04:47.903904Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"nCUjOkuGECr3j0M1bMIse5cYOjNBx3geVGkTQ9paZqKPc6s82qZL9eM0jJhrdZmdRrt2OONVuDdMhf2Te80eBQ==","signature_status":"signed_v1","signed_at":"2026-05-28T02:04:47.904341Z","signed_message":"canonical_sha256_bytes"},"source_id":"2604.16358","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:19cf3f55b2b1f545b8c3163f7db1c847dd0218f0d3d1ab33c56debefbb9071aa","sha256:c8158fd9e84cc92d6edc49fbe9d5b323580bb100649e6efdc00ca8b27730083d"],"state_sha256":"047c6c6ae5b079039e77f16ffd5b677fc1e7d02bc98cb4be4a00025627db39d0"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"K2m93OrYhCYq98Xg6ZN4HpfXHcGkauM4J/yE/AuGr67xC5KQBZDZD5BsIquShYLO8t2xhePK8eEFNTAcO/MMDw==","signed_message":"bundle_sha256_bytes","signed_at":"2026-06-08T16:46:06.081256Z","bundle_sha256":"da953d1e992d0171b7da103665df4e8fcf126357292ca073b51269716a6af609"}}