Research map / Vision-language and multimodal / Multimodal LLMs
Two Birds, One Projection: Harmonizing Safety and Utility in LVLMs via Inference-time Feature Projection
Yewon Han, Yumin Seol, Minsoo Jo, EunGyung Kong, Taesup Kim
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 |
|---|---|---|---|
| apply TBOP null-space projection | LLaVA-1.6-7B · MM-SafetyBench SD-TYPO | ASR (%): 38.86 -> 5.09 | Improved |
| apply TBOP null-space projection | LLaVA-1.6-7B · MME | score: 1715.13 -> 1870.17 | Improved |
| apply TBOP null-space projection | LLaVA-1.6-7B · ScienceQA | accuracy: 57.20 -> 63.10 | Improved |
| apply TBOP null-space projection | LLaVA-1.6-13B · MM-Vet | score: 45.63 -> 49.98 | Improved |
| apply TBOP null-space projection | LLaVA-1.6-7B · HADES | ASR (%): 39.06 -> 5.48 | Improved |
| apply detect-then-detoxify defenses such as Immune or Protector instead of TBOP | LLaVA-1.6-7B · MME | score: 1715.13 -> 725.39 | Hurt |
| add TBOP on top of ECSO | ECSO · FigStep with SI-Attack (7B) | ASR (%): 16.8 -> 5.6 | Improved |
| add TBOP on top of ETA (13B) | ETA · FigStep with SI-Attack (13B) | ASR (%): 0.8 -> 0.6 | Improved |
| use MM-SafetyBench-only anchor instead of mixed anchor (ablation) | TBOP · MM-SafetyBench SD-TYPO | ASR (%): 5.09 -> 7.10 | Hurt |
| use MME-only anchor instead of mixed anchor (ablation) | TBOP · MM-SafetyBench SD-TYPO | ASR (%): 5.09 -> 36.54 | Hurt |
| use a blank dummy image instead of Gaussian noise (ablation) | TBOP · MM-SafetyBench SD-TYPO | ASR (%): 10.12 -> 5.09 | Improved |
| use a blank dummy image instead of uniform noise (ablation) | TBOP · MM-SafetyBench SD-TYPO | ASR (%): 10.26 -> 5.09 | 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.
- Immune — Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment (CVPR 2025) · better on 2 of 2 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
- ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time (ICLR 2025 · similarity 0.93)
- JailBound: Jailbreaking Internal Safety Boundaries of Vision-Language Models (NeurIPS 2025 · similarity 0.91)
- Diagnosing and Repairing Unsafe Channels in Vision-Language Models via Causal Discovery and Dual-Modal Safety Subspace Projection (CVPR 2026 · similarity 0.90)
- Safe + Safe = Unsafe? Exploring How Safe Images Can Be Exploited to Jailbreak Large Vision-Language Models (NeurIPS 2025 · similarity 0.90)
- Cross-Modal Safety Mechanism Transfer in Large Vision-Language Models (ICLR 2025 · similarity 0.90)
- Risk Awareness Injection: Calibrating Vision-Language Models for Safety without Compromising Utility (ICML 2026 · similarity 0.90)
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.
