Skip to content

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

Oct 12 · Monday

All papers

Found 13 papers

Another 262papers do not have category tags yet and are not included in filtered results.

  1. DefensesarXiv 2610.01170

    HeadEdit:通过冻结的 unembedding 矩阵校准语言模型行为

    提出 HeadEdit,经冻结 unembedding 做随输入变化的 logit 校正

    Published
    Evaluated models
    Qwen3-4B, Gemma-3-4B-IT, Llama-3.2-3B-Instruct and others4
    Abstract低秩 unembedding 校正抬高九个设置的总体准确率,并可部分预测 DPO 更新。

    HeadEdit 是一种不更新参数的行为校准方法,作用点是冻结的 unembedding,而不是中间层残差或新的梯度优化。

    Paper analysisRead analysis
  2. AgentBound:面向工具型 LLM Agent 安全的四路反事实评测框架

    提出 AGENT BOUNDARY,用四向反事实任务评测工具智能体行动边界

    Published
    Setting
    AI Agent
    Evaluated models
    GPT-5.5, Claude Opus 4.8, Gemini 3.5 Flash and others15
    Abstract四向反事实基准把过度拒绝和未授权遵从拆开,Thought 校正在十个配置上提高已授权完成。

    AGENT BOUNDARY 是面向工具使用 LLM 智能体的四向可执行反事实评测,并附带一个决策时 Thought 校准模块。

    Paper analysisRead analysis
  3. DefensesarXiv 2609.33086

    研究者提出基于评分量表的可解释奖励框架,将宪法转化为可调对齐目标

    提出可解释 rubric 奖励,将宪法转为可调权重并用于监督微调

    Published
    Evaluated models
    Gemma 3 12B, Qwen3 14B, GPT-OSS 20B
    Abstract宪法级可解释奖励让优先级可检查、可改写,并带入训练后的行为。

    作者给出一条从通用宪法到可调训练奖励的路径,用来检查并改写模型实际被优化的优先级。

    Paper analysisRead analysis
  4. Knowing What Not to Answer: Selective Non-Compliance in Vision-Language Models

    Propose the KoNA benchmark to evaluate selective non-compliance in vision-language models

    Published
    Evaluated models
    InternVL3-2B-Instruct, InternVL3-78B-Instruct, Qwen2.5-VL-3B-Instruct and others8
    AbstractKoNA evaluates selective noncompliance with five types of paired queries.

    KoNA is used to evaluate whether vision-language models can selectively refuse mixed queries: only handle the components that should not be followed, while answering the remaining components.

    Paper analysisRead analysis
  5. DefensesarXiv 2609.04714

    Refusal without refusal: an analysis of safety fine-tuning response structure for reducing language model false refusals

    Proposes Rationale-Only safety fine-tuning, removing the refusal statement to reduce false refusals.

    Published
    Evaluated models
    Llama-3.1-8B, Mistral-7B-v0.3, Gemma-2-9B and others8
    AbstractRemoving the boilerplate refusal statement and training only on the rationale can reduce false refusals, while the compliance rate on harmful queries remains low.

    This is a structural analysis of safety fine-tuning data: false refusals are traced by the authors to formulaic refusal statements in responses, rather than relying only on inference-time remedies.

    Paper analysisRead analysis
  6. Study audits 9 vision-language models: national stance rewriting rate under Chinese prompts is about three times that under English

    审计视觉语言模型在敏感图像上从拒答转向国家立场重构的行为

    Published
    Evaluated models
    Qwen2-VL-7B, Qwen2.5-VL-7B, Qwen3-VL-8B and others9
    Abstract论文系统审计了 VLM 多模态审查,发现审查正从显式拒答转向隐蔽重构,且中文提示与文本先验会放大重构倾向。

    论文系统研究了中外开源视觉语言模型在政治敏感图像上的审查行为,主要内容如下。

    Paper analysisRead analysis
  7. BioSecBench-Refusal:面向智能体生物安全风险评估的配对基准

    提出 BioSecBench-Refusal,评测智能体生物研发拒答与隐蔽危害识别

    Published
    Evaluated models
    Opus 4.8, Opus 4.7, Opus 4.6 and others10
    Abstract配对基准显示拒绝多由 API 触发,合法任务常被拒得不少于隐蔽危害。

    BioSecBench-Refusal 把生物研发智能体的谨慎程度和识别隐蔽危害的能力放在同一套任务上测量。

    Paper analysisRead analysis
  8. 研究用 30 项漏洞分析任务评测自主安全智能体的安全对齐效果

    评测自主安全智能体在授权漏洞分析中的对齐拒答效应

    Published
    Evaluated models
    Gemma 4 31B official instruction-tuned model, TrevorJS Gemma 4 31B uncensored GGUF, Gemma 4 26B A4B official instruction-tuned model and others8
    Abstract轨迹级评测把拒答、接地、工具接口分开。

    这项工作把安全对齐效应测成自主安全智能体的轨迹属性,而不是单轮拒答率。

    Paper analysisRead analysis
  9. 分析大语言模型在仇恨言论去毒中误拒行为的偏差

    分析 LLM 仇恨言论解毒误拒的群体偏差并测试跨语翻译缓解

    Published
    Evaluated models
    GPT-4o mini, GPT-3.5 turbo, Llama-3.1 8B and others9
    Abstract九个 LLM 在解毒任务上普遍误拒,国籍宗教政治类偏高。

    作者考察对齐后的 LLM 是否会拒绝可接受的仇恨言论解毒请求,并分析误拒是否随目标群体和语言变化。完全拒答与夹带说教的部分拒答都计入。设置是固定提示、九个 LLM、三个英文集和法德西中韩。

    Paper analysisRead analysis
You have reached the end