mlsanityData to 14 Oct 2026

ReShift: Aha-Moment-Driven Reasoning-Level Backdoor Attacks on Vision-Language Models

arXiv 2607.00361 ↗First posted 1 Jul 2026cs.CRcited by 0

On the map: LLM safety and security → Backdoor and data poisoning attacks

Abstract (arXiv)

Vision--Language Models (VLMs) are increasingly deployed in safety-critical applications, yet remain vulnerable to backdoor attacks. Existing methods primarily manipulate final outputs, often producing reasoning traces that are inconsistent or easily detectable. In this paper, we propose ReShift, the novel aha-moment-driven reasoning-level backdoor framework that explicitly redirects the internal chain-of-thought (CoT) trajectory while preserving surface-level coherence. ReShift introduces a Poisoned Reasoning-Aware Data Construction (PRDC) pipeline and a Supervised--Reinforcement Joint Optimization (SRJO) strategy to induce stable trigger-conditioned reasoning shifts. We further formalize Entropy Rebound as a principled signal for characterizing reasoning redirection and provide theoretical guaranties linking entropy gaps to trajectory-level divergence. Extensive experiments demonstrate that ReShift achieves high attack success rates while maintaining clean-task performance and realistic reasoning traces, substantially improving stealthiness against existing defenses.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Backdoor and data poisoning attacks, one of 12 directions in LLM safety and security. Builds on 48 of its 55 references that are in the map. First posted on arXiv 1 Jul 2026.

Builds on

48 of 55 references.

Closest ancestors · 13

Same direction, newest first

  1. 2025TokenSwap: Backdoor Attack on the Compositional Understanding of Large Vision-Language Models29 Sep
  2. 2025Backdoor Attribution: Elucidating and Controlling Backdoor in Language Models26 Sep
  3. 2025Backdoor Cleaning without External Guidance in MLLM Fine-tuning22 May
  4. 2025ShadowCoT: Cognitive Hijacking for Stealthy Reasoning Backdoors in LLMs8 Apr
  5. 2025BadToken: Token-level Backdoor Attacks to Multi-modal Large Language Models20 Mar
  6. 2025Stealthy Backdoor Attack in Self-Supervised Learning Vision Encoders for Large Vision Language Models25 Feb
  7. 2024Backdooring Vision-Language Models with Out-Of-Distribution Data2 Oct
  8. 2024TrojVLM: Backdoor Attack Against Vision Language Models28 Sep
  9. 2024Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift27 Jun
  10. 2024VL-Trojan: Multimodal Instruction Backdoor Attacks against Autoregressive Visual Language Models21 Feb
  11. 2024Test-Time Backdoor Attacks on Multimodal Large Language Models13 Feb
  12. 2023BadCLIP: Trigger-Aware Prompt Learning for Backdoor Attacks on CLIP26 Nov
1 more
  1. 2021Dual-Key Multimodal Backdoors for Visual Question Answering14 Dec

Other references · 35

  1. 2017Proximal Policy Optimization Algorithmscited by 15,215
  2. 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language Modelscited by 12,459
  3. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 9,379
  4. 2022ReAct: Synergizing Reasoning and Acting in Language Modelscited by 8,695
  5. 2022Self-Consistency Improves Chain of Thought Reasoning in Language Modelscited by 5,708
  6. 2025Qwen2.5-VL Technical Reportcited by 5,587
  7. 2023Improved Baselines with Visual Instruction Tuningcited by 4,418
  8. 2023Tree of Thoughts: Deliberate Problem Solving with Large Language Modelscited by 3,211
  9. 2023InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuningcited by 2,876
  10. 2025DAPO: An Open-Source LLM Reinforcement Learning System at Scalecited by 2,810
  11. 2023MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGIcited by 2,080
  12. 2023PaLM-E: An Embodied Multimodal Language Modelcited by 2,071
23 more
  1. 2022Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answeringcited by 1,919
  2. 2023MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contextscited by 1,614
  3. 2025InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiencycited by 1,510
  4. 2023Multimodal Chain-of-Thought Reasoning in Language Modelscited by 752
  5. 2022A-OKVQA: A Benchmark for Visual Question Answering using World Knowledgecited by 710
  6. 2023MiniGPT-v2: large language model as a unified interface for vision-language multi-task learningcited by 625
  7. 2024Measuring Multimodal Mathematical Reasoning with MATH-Vision Datasetcited by 526
  8. 2025Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRscited by 377
  9. 2025VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Models with Reinforcement Learningcited by 323
  10. 2023Vision Language Models in Autonomous Driving: A Survey and Outlookcited by 126
  11. 2023Visual Chain of Thought: Bridging Logical Gaps with Multimodal Infillingscited by 69
  12. 2025SRPO: Enhancing Multimodal LLM Reasoning via Reflection-Aware Reinforcement Learningcited by 57
  13. 2024MC-Bench: A Benchmark for Multi-Context Visual Grounding in the Era of MLLMscited by 39
  14. 2024Physical Backdoor Attack can Jeopardize Driving with Vision-Large-Language Modelscited by 33
  15. 2025Painting with Words: Elevating Detailed Image Captioning with Benchmark and Alignment Learningcited by 25
  16. 2025Embodied Scene Understanding for Vision Language Models via MetaVQAcited by 19
  17. 2025Plan Then Action:High-Level Planning Guidance Reinforcement Learning for LLM Reasoningcited by 19
  18. 2024Guiding Vision-Language Model Selection for Visual Question-Answering Across Tasks, Domains, and Knowledge Typescited by 10
  19. 2026STRIDE: Strategic Trajectory Reasoning via Discriminative Estimation for Verifiable Reinforcement Learningcited by 10
  20. 2026The Illusion of Insight in Reasoning Modelscited by 9
  21. 2025DSADF: Thinking Fast and Slow for Decision Makingcited by 7
  22. 2025CARE-PD: A Multi-Site Anonymized Clinical Dataset for Parkinson's Disease Gait Assessmentcited by 6
  23. 2025Trust-Region Adaptive Policy Optimizationcited by 3

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026STAR: Detecting Inference-time Backdoors in LLM Reasoning via State-Transition Amplification Ratio
  2. 2026Thinking Wrong in Silence: Backdoor Attacks on Continuous Latent Reasoning
  3. 2025BackdoorVLM: A Benchmark for Backdoor Attacks and Defenses on Vision-Language Models
  4. 2026Detecting Chain-of-Thought Hijacking in Backdoored Multimodal Large Reasoning Models
  5. 2026Phantasia: Context-Adaptive Backdoors in Vision Language Models
  6. 2026Critical-CoT: A Robust Defense Framework against Reasoning-Level Backdoor Attacks in Large Language Models
  7. 2025BadReasoner: Planting Tunable Overthinking Backdoors into Large Reasoning Models for Fun or Profit
  8. 2026Reasoning-Oriented Programming: Chaining Semantic Gadgets to Jailbreak Large Vision Language Models