Skip to content

Translation in progress Some entries still include text in the original language.

Oct 12 · Monday

All papers

Found 3 papers
Filter4

Model risks

Deception and scheming10Reward hacking3Misalignment and value divergence3Inappropriate refusal7Hallucination and fabrication11Bias and discrimination3

Layer

Model layer3

System component

Reasoning chains2Multimodal1
Select any value within a category; different categories must all match. Counts include the other filters you have selected.Another 240papers have no tags yet and are not included in filtered results.
Model risks：Misalignment and value divergenceLayer：Model layerSystem component：Reasoning chainsSystem component：MultimodalClear all

Another 240papers do not have category tags yet and are not included in filtered results.

  1. Risky behaviorarXiv 2609.35291

    窄域多模态微调可诱发涌现性失配，覆盖 15 个视觉语言模型

    证明无显式危害的图文窄微调可诱发跨任务涌现未对齐

    arXiv
    2609.35291
    Published
    Layer
    Model layer
    Setting
    Browser and computer use
    Evaluated models
    GPT-4o, GPT-4.1, Gemini 2.5 and others15
    Abstract证实窄多模态微调可引发全面的多渠道未对齐行为转变，揭示了后训练阶段对齐被隐蔽重塑的安全隐患。

    论文系统探究了视觉-语言模型中由窄任务微调诱发的跨模态涌现未对齐（Emergent Misalignment, EM）现象。

    Paper analysisRead analysis
  2. Risky behaviorarXiv 2609.08186

    论文揭示大推理模型的对齐崩溃并提出 Reasoning Trap 越狱范式

    揭示推理加深引发的对齐坍塌，用 Reasoning Trap 放大越狱

    arXiv
    2609.08186
    Published
    Layer
    Model layer
    Setting
    Models and APIs
    Evaluated models
    Qwen3-8B, Qwen3-14B, Llama2-7B and others8
    Abstract更深推理提升题目准确率，同时提高扰动下的相对损失。

    作者研究扩展推理是否以对齐稳健性为代价，提出 ALR、RT 与 RRA。。

    Paper analysisRead analysis
  3. Risky behaviorarXiv 2607.14345

    Truthful AI 提出价值泄漏评估：模型答案被自身价值观悄然左右

    提出价值泄漏评估，测量模型回答受自身价值倾向的隐蔽影响

    arXiv
    2607.14345
    Published
    Layer
    Model layer
    Setting
    Models and APIs
    Evaluated models
    Claude Opus 4.8, Claude Opus 4.7, Claude Opus 4.6 and others15

    Abstract论文揭示前沿模型回答常受自身价值隐蔽塑造，提出反事实评估量化偏见及其思维链隐瞒，指出当前对齐未能有效防范该失效。

    Paper analysisRead analysis
You have reached the end