Research map / Large language models / Alignment and preferences
Meerkat-VL: Implicit Risk Safety Alignment in Multimodal LLMs via Perceptual Reasoning and Self-Verification
Peicheng Zhou, Chuanbin Liu, Shancheng Fang, Bowei Pu, Yiwei Sun, Zhangchi Hu, Hongtao Xie
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 |
|---|---|---|---|
| add perceptual reward and consistency penalty (DPCA) together (ablation) | Meerkat-VL on Qwen2.5-VL-7B · SIUO safety | safety score (%): 73.05 -> 89.68 | Improved |
| add perceptual reward and consistency penalty (DPCA) together (ablation) | Meerkat-VL on Qwen2.5-VL-7B · MSSBench safety | safety score (%): 72.33 -> 88.33 | Improved |
| add consistency penalty on top of perceptual reward (ablation) | Meerkat-VL · SIUO safety | safety score (%): 83.23 -> 89.68 | Improved |
| add perceptual reward on top of consistency penalty (ablation) | Meerkat-VL · SIUO safety | safety score (%): 86.23 -> 89.68 | Improved |
| perceptual reward without consistency penalty (ablation) | Meerkat-VL · explicit risk average | helpfulness (%): 93.63 -> 90.09 | Hurt |
| replace CM-V/RM-V reward models with NPSV self-verification (ablation) | Qwen2.5-VL-7B · reward accuracy helpfulness | accuracy (%): 77.1 -> 92.2 | Improved |
| replace CM-V/RM-V reward models with NPSV on Beavertails-V (ablation) | Qwen2.5-VL-7B · SIUO safety | safety score (%): 52.3 -> 70.7 | Improved |
| train on Meerkat-Safe instead of Beavertails-V with NPSV (ablation) | Qwen2.5-VL-7B · SIUO safety | safety score (%): 70.7 -> 89.7 | Improved |
| train on SPA-VL preference data | Qwen2.5-VL-7B · SIUO safety | safety score (%): 37.13 -> 39.52 | No clear effect |
| train on SPA-VL preference data | Qwen2.5-VL-7B · average helpfulness | helpfulness (%): 73.72 -> 61.90 | Hurt |
| replace conventional GRPO reward models with NPSV | Meerkat-VL · training cost | total training hours: 8.27h -> 6.87h (-16.93%) | Improved |
| add perceptual reasoning tokens | Meerkat-VL · response length | tokens: 382.0 / 385.0 -> 397.0 / 384.0 | No clear effect |
Closest papers by meaning
- Pragma-VL: Towards a Pragmatic Arbitration of Safety and Helpfulness in MLLMs (ICLR 2026 · similarity 0.90)
- Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedback (NeurIPS 2025 · similarity 0.89)
- SafeGRPO: Self-Rewarded Multimodal Safety Alignment via Rule-Governed Policy Optimization (CVPR 2026 · similarity 0.89)
- Safe Delta: Consistently Preserving Safety when Fine-Tuning LLMs on Diverse Datasets (ICML 2025 · similarity 0.88)
- Risk Awareness Injection: Calibrating Vision-Language Models for Safety without Compromising Utility (ICML 2026 · similarity 0.87)
- Safety Geometry Collapse in Multimodal LLMs and Adaptive Drift Correction (NeurIPS 2026 · similarity 0.87)
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.
