mlsanityData to 7 Oct 2026

Towards Multi-modal Multi-turn Safety: From Agentic Interaction to Strategic Alignment

arXiv 2601.04736 ↗First posted 8 Jan 2026cs.CLcited by 0

On the map: Multimodal LLMs / vision-language models → MLLM jailbreaks and adversarial robustness

also in 2 more
Abstract (arXiv)

Despite remarkable capability in multi-modal understanding, deploying Multi-modal Large Language Models (MLLMs) in open-ended conversational scenarios introduces safety risks that remain poorly addressed by existing alignment methods. Unlike simple malicious visual question and answer (VQA) pairs , multi-turn interactions enable adversaries to incrementally reconstruct harmful intent across dialogues, progressively bypassing safety constraints in ways that are difficult to detect at any individual turn. Meanwhile, conventional reinforcement learning from human feedback (RLHF) approaches are unsuitable for this situation: designed primarily for VQA tasks, they neither capture cross-turn risk dynamics nor scale efficiently without costly manual preference annotation. To close this gap, we introduce MINT-Safe, an open-source visual multi-turn training dataset comprising 11,270 multi-image dialogues and 500 refusal VQA pairs, constructed via multi-agent interaction with text-to-image (T2I) tool-call augmentation. Building on MINT-Safe, we propose TAD-Align, a dialogue safety alignment framework centered on a turn-aware dual-objective reward function. Rather than treating all dialogue turns uniformly, TAD-Align leverages rollout-based safety score variance to dynamically identify turns where the model exhibits inconsistent safety behavior, and adaptively up-weights these turns during optimization. Experiments on Qwen2.5-VL-7B-Instruct and LLaVA-NeXT-7B demonstrate reductions of over 10% in Attack Success Rate (ASR), alongside improvements of at least 8% in harmlessness and 13% in helpfulness on multi-modal multi-turn safety benchmarks, while preserving general model capabilities.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in MLLM jailbreaks and adversarial robustness, one of 16 directions in Multimodal LLMs / vision-language models. Builds on 28 of its 49 references that are in the map. First posted on arXiv 8 Jan 2026.

Builds on

28 of 49 references.

Closest ancestors · 5

Same direction, newest first

  1. 2025SafeGRPO: Self-Rewarded Multimodal Safety Alignment via Rule-Governed Policy Optimization17 Nov
  2. 2025SafeMT: Multi-turn Safety for Multimodal Language Models14 Oct
  3. 2025Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedback22 Mar
  4. 2024VLSBench: Unveiling Visual Leakage in Multimodal Safety29 Nov
  5. 2024JailBreakV: A Benchmark for Assessing the Robustness of MultiModal Large Language Models against Jailbreak Attacks3 Apr

Other references · 23

  1. 2017Proximal Policy Optimization Algorithmscited by 14,441
  2. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 8,851
  3. 2023Judging LLM-as-a-Judge with MT-Bench and Chatbot Arenacited by 8,393
  4. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 7,916
  5. 2023MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contextscited by 1,582
  6. 2023Aligning Large Multimodal Models with Factually Augmented RLHFcited by 703
  7. 2021Inter-GPS: Interpretable Geometry Problem Solving with Formal Language and Symbolic Reasoningcited by 376
  8. 2024PKU-SafeRLHF: Towards Multi-Level Safety Alignment for LLMs with Human Preferencecited by 252
  9. 2024Attacks, Defenses and Evaluations for LLM Conversation Safety: A Surveycited by 124
  10. 2025X-Teaming: Multi-Turn Jailbreaks and Defenses with Adaptive Multi-Agentscited by 112
  11. 2023Are You Sure? Challenging LLMs Leads to Performance Drops in The FlipFlop Experimentcited by 71
  12. 2025PrivaCI-Bench: Evaluating Privacy with Contextual Integrity and Legal Compliancecited by 34
11 more
  1. 2025Mitigating Deceptive Alignment via Self-Monitoringcited by 25
  2. 2025Steering Dialogue Dynamics for Robustness against Multi-turn Jailbreaking Attackscited by 20
  3. 2025ThinkPatterns-21k: A Systematic Study on the Impact of Thinking Patterns in LLMscited by 16
  4. 2025Leveraging LLMs as Meta-Judges: A Multi-Agent Framework for Evaluating LLM Judgmentscited by 16
  5. 2025Generative RLHF-V: Learning Principles from Multi-modal Human Preferencecited by 13
  6. 2025MinD: Learning A Dual-System World Model for Real-Time Planning and Implicit Risk Analysiscited by 13
  7. 2025Firm or Fickle? Evaluating Large Language Models Consistency in Sequential Interactionscited by 12
  8. 2025GSPR: Aligning LLM Safeguards as Generalizable Safety Policy Reasonerscited by 10
  9. 2025SafeLawBench: Towards Safe Alignment of Large Language Modelscited by 9
  10. 2024Systematic Reward Gap Optimization for Mitigating VLM Hallucinationscited by 7
  11. 2025InterMT: Multi-Turn Interleaved Preference Alignment with Human Feedbackcited by 3

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025LLaVAShield: Safeguarding Multimodal Multi-Turn Dialogues in Vision-Language Models
  2. 2026One Turn Too Late: Learning When to Intervene Against Multi-Turn Malicious Intent
  3. 2026A2Safe: Counterfactual Evidence-Aligned Adaptive Agent Collaboration for Safe and Effective Visual Question Answering
  4. 2026MTMCS-Bench: Evaluating Contextual Safety of Multimodal Large Language Models in Multi-Turn Dialogues
  5. 2025SafeTy Reasoning Elicitation Alignment for Multi-Turn Dialogues
  6. 2026ReFrame: Evidence-Guided Test-Time Safety Alignment in Multimodal Large Language Models
  7. 2024Cross-Modal Safety Alignment: Is textual unlearning all you need?
  8. 2026Learning When to Act or Refuse: Guarding Agentic Reasoning Models for Safe Multi-Step Tool Use