AI Paper Atlas

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

ICML 2026 · Large language models › Alignment and preferences · approach: Steering and correcting LLM behavior

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 changedWhereMeasuredEffect
add perceptual reward and consistency penalty (DPCA) together (ablation)Meerkat-VL on Qwen2.5-VL-7B · SIUO safetysafety score (%): 73.05 -> 89.68Improved
add perceptual reward and consistency penalty (DPCA) together (ablation)Meerkat-VL on Qwen2.5-VL-7B · MSSBench safetysafety score (%): 72.33 -> 88.33Improved
add consistency penalty on top of perceptual reward (ablation)Meerkat-VL · SIUO safetysafety score (%): 83.23 -> 89.68Improved
add perceptual reward on top of consistency penalty (ablation)Meerkat-VL · SIUO safetysafety score (%): 86.23 -> 89.68Improved
perceptual reward without consistency penalty (ablation)Meerkat-VL · explicit risk averagehelpfulness (%): 93.63 -> 90.09Hurt
replace CM-V/RM-V reward models with NPSV self-verification (ablation)Qwen2.5-VL-7B · reward accuracy helpfulnessaccuracy (%): 77.1 -> 92.2Improved
replace CM-V/RM-V reward models with NPSV on Beavertails-V (ablation)Qwen2.5-VL-7B · SIUO safetysafety score (%): 52.3 -> 70.7Improved
train on Meerkat-Safe instead of Beavertails-V with NPSV (ablation)Qwen2.5-VL-7B · SIUO safetysafety score (%): 70.7 -> 89.7Improved
train on SPA-VL preference dataQwen2.5-VL-7B · SIUO safetysafety score (%): 37.13 -> 39.52No clear effect
train on SPA-VL preference dataQwen2.5-VL-7B · average helpfulnesshelpfulness (%): 73.72 -> 61.90Hurt
replace conventional GRPO reward models with NPSVMeerkat-VL · training costtotal training hours: 8.27h -> 6.87h (-16.93%)Improved
add perceptual reasoning tokensMeerkat-VL · response lengthtokens: 382.0 / 385.0 -> 397.0 / 384.0No clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy