Papers5

#out-of-distribution

When the Prompt Becomes Visual: Vision-Centric Jailbreak Attacks for Large Image Editing Models

Jiacheng Hou, Yining Sun et al.Feb 10arXiv

Modern image editors can now follow visual prompts like arrows and scribbles, which opens a new way for attackers to hide harmful instructions inside images.

#vision-centric jailbreak#image editing safety#visual prompts

Effective Reasoning Chains Reduce Intrinsic Dimensionality

Beginner

Archiki Prasad, Mandar Joshi et al.Feb 9arXiv

The paper asks a simple question: which kind of step-by-step reasoning helps small language models learn best, and why?

#intrinsic dimensionality#chain-of-thought#LoRA

RLAnything: Forge Environment, Policy, and Reward Model in Completely Dynamic RL System

Beginner

Yinjie Wang, Tianbao Xie et al.Feb 2arXiv

RLAnything is a new reinforcement learning (RL) framework that trains three things together at once: the policy (the agent), the reward model (the judge), and the environment (the tasks).

#reinforcement learning#closed-loop optimization#reward modeling

Thinking in Frames: How Visual Context and Test-Time Scaling Empower Video Reasoning

Intermediate

Chengzu Li, Zanyi Wang et al.Jan 28arXiv

This paper shows that making short videos can help AI plan and reason in pictures better than writing out steps in text.

#video reasoning#visual planning#test-time scaling

OmniSafeBench-MM: A Unified Benchmark and Toolbox for Multimodal Jailbreak Attack-Defense Evaluation

Intermediate

Xiaojun Jia, Jie Liao et al.Dec 6arXiv

OmniSafeBench-MM is a one-stop, open-source test bench that fairly compares how multimodal AI models get tricked (jailbroken) and how well different defenses stop that.

#multimodal large language models#jailbreak attacks#safety alignment