Research map / Adversarial robustness and security / Attacks on text and graphs
ReShift: Aha-Moment-Driven Reasoning-Level Backdoor Attacks on Vision–Language Models
Zhihao Dou, Qinjian Zhao, Zhiqiang Gao, Sumon Biswas
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| redirect reasoning trajectory (ReShift) instead of overriding output (BadToken) | Qwen2.5-VL-7B · A-OKVQA | ASR: 0.92 -> 0.97 | Improved |
| use ReShift versus Rewrite on out-of-domain data | Qwen2.5-VL-7B · MMMU | ASR: 0.55 -> 0.74 | Improved |
| add target reward R_target (ablation) | ReShift · A-OKVQA | ASR: 0.48 -> 0.97 | Improved |
| add shift reward R_shift (ablation) | ReShift · A-OKVQA | coherence: 3.35 -> 4.03 | Improved |
| add shift reward R_shift (ablation) | ReShift · A-OKVQA | ASR: 0.96 -> 0.97 | No clear effect |
| add format reward R_format (ablation) | ReShift · A-OKVQA | coherence: 3.86 -> 4.03 | Improved |
| add RL on suffix to SFT-only injection (ablation) | ReShift · A-OKVQA | coherence: 3.27 -> 4.03 | Improved |
| add RL on suffix to SFT-only injection (ablation) | ReShift · A-OKVQA | ASR: 0.93 -> 0.97 | Improved |
| use reasoning-level attack against detector BYE | ReShift · A-OKVQA | DACC: 0.72 -> 0.11 | Improved |
| increase training data from 2000 to 6600 samples (ablation) | ReShift · A-OKVQA | ASR: 0.92 -> 0.97 | Improved |
| increase training data from 2000 to 6600 samples (ablation) | ReShift · A-OKVQA | clean accuracy: 0.78 -> 0.87 | Improved |
| increase GRPO group size from 1 to 4 (ablation) | ReShift · four benchmarks | ASR | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- BadToken — BadToken: Token-level Backdoor Attacks to Multi-modal Large Language Models (CVPR 2025) · better on 7 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · similarity 0.89)
- Semantic Shield: Defending Vision-Language Models Against Backdooring and Poisoning via Fine-grained Knowledge Alignment (CVPR 2024 · similarity 0.86)
- BadVLA: Towards Backdoor Attacks on Vision-Language-Action Models via Objective-Decoupled Optimization (NeurIPS 2025 · similarity 0.86)
- Anchoring the Mind of Multimodal Reasoners: Cognitive Bias as a Vector for Jailbreak Attacks (CVPR 2026 · similarity 0.86)
- BadThink: Triggered Overthinking Attacks on Chain-of-Thought Reasoning in Large Language Models (AAAI 2026 · similarity 0.86)
- Steering Away from Harm: An Adaptive Approach to Defending Vision Language Model Against Jailbreaks (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
