Research map / Vision-language and multimodal / Multimodal LLMs
Attention Misses Visual Risk: Risk-Adaptive Steering for Multimodal Safety Alignment
Jonghyun Park, Minhyuk Seo, Chaewon YEO, Jonghyun Choi
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 MoRAS (visual context + risk-scaled steering) | LLaVA-1.5-7B · SPA-VL | ASR: 47.2 -> 7.6 | Improved |
| apply MoRAS | LLaVA-1.5-7B · FigStep | ASR: 59.3 -> 2.8 | Improved |
| apply MoRAS | LLaVA-1.5-7B · VAA | ASR: 43.1 -> 14.3 | Improved |
| apply MoRAS | LLaVA-1.5-7B · MM-Vet | score: 30.5 -> 30.5 | No clear effect |
| apply MoRAS | LLaVA-OneVision-7B · MM-Vet | score: 52.8 -> 50.8 | Hurt |
| apply MoRAS | Qwen-VL-Chat · MM-Vet | score: 48.7 -> 46.9 | Hurt |
| add risk-adaptive steering (stages 2-3) on top of Stage 1 reformulation (ablation) | MoRAS · MM-Vet | score: 28.6 -> 30.5 | Improved |
| add risk-adaptive steering (stages 2-3) on top of Stage 1 (ablation) | MoRAS · SPA-VL | ASR: 7.1 -> 7.6 | No clear effect |
| use query-to-unsafe-prototype refusal vector instead of safe/unsafe mean difference (last layer) (ablation) | MoRAS · LLaVA-1.5-7B avg ASR | ASR: 25.9 -> 8.3 | Improved |
| steer last layer instead of layer 14 (ablation) | MoRAS · LLaVA-1.5-7B avg ASR | ASR: 13.1 -> 8.3 | Improved |
| calibrate with full dataset instead of 100 samples (ablation) | MoRAS · SPA-VL | ASR: 7.6 -> 7.9 (Wu et al.) | No clear effect |
| change unsafe-query source for prototypes (GPT-4 vs LLM-RTD vs I-Malicious) (ablation) | MoRAS · SPA-VL | ASR: 7.6 / 6.4 / 5.7 | No clear effect |
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.
- CoCA — COCA: COllaborative CAusal Regularization for Audio-Visual Question Answering (AAAI 2023) · better on 1 of 1 shared result
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
- VLSU: Mapping the Limits of Joint Multimodal Understanding for AI Safety (ICLR 2026 · similarity 0.89)
- Pragma-VL: Towards a Pragmatic Arbitration of Safety and Helpfulness in MLLMs (ICLR 2026 · similarity 0.89)
- ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time (ICLR 2025 · similarity 0.89)
- GuardAlign: Robust Safety Alignment in Multimodal Large Language Models (ICLR 2026 · similarity 0.89)
- SafeGRPO: Self-Rewarded Multimodal Safety Alignment via Rule-Governed Policy Optimization (CVPR 2026 · similarity 0.88)
- Rethinking Bottlenecks in Safety Fine-Tuning of Vision Language Models (ICLR 2026 · similarity 0.88)
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.
