<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[Conger_2002的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/Conger_2002</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; Conger_2002]]></copyright><item><title><![CDATA[Positional Overload: Positional Debiasing and Context Window Extension for LLM using Set Encoding]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/155647798</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/155647798</guid><author>Conger_2002</author><pubDate>Sun, 07 Dec 2025 09:09:30 +0800</pubDate><description><![CDATA[集合编码：让模型将一个集合（一组无序的元素）视为一个整体，其内部元素的排列顺序不会影响模型的输出。]]></description><category></category></item><item><title><![CDATA[Large Language Models Sensitivity to The Order of Options in Multiple-Choice Questions 2024.finding]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/155647752</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/155647752</guid><author>Conger_2002</author><pubDate>Sat, 06 Dec 2025 21:31:51 +0800</pubDate><description><![CDATA[LLM 对选项顺序的敏感程度有多大？]]></description><category></category></item><item><title><![CDATA[Order Matters: Investigate the Position Bias in Multi-constraint Instruction Following 2025.findings]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/155647737</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/155647737</guid><author>Conger_2002</author><pubDate>Sat, 06 Dec 2025 21:27:47 +0800</pubDate><description><![CDATA[构建大量多约束指令，改变约束顺序并计算CDDI值，发现大模型在“难到易”顺序下表现最好，并通过注意力可视化解释了原因：模型在难到易顺序下更关注约束内容。]]></description><category></category></item><item><title><![CDATA[Option Symbol Matters: Investigating and Mitigating Multiple-Choice Option Symbol Bias of LLM 25ACL]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/155647682</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/155647682</guid><author>Conger_2002</author><pubDate>Sat, 06 Dec 2025 21:24:40 +0800</pubDate><description><![CDATA[选项符号偏差（Option Symbol Bias） A/B/C/D） i-iv）1-4）x₁–x₄），是由于关键注意力头过度关注选项符号本身，而非其语义内容；与“下一词预测”预训练范式有关]]></description><category></category></item><item><title><![CDATA[LARGE LANGUAGE MODELS ARE NOT ROBUST ICLR2024]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/155647581</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/155647581</guid><author>Conger_2002</author><pubDate>Sat, 06 Dec 2025 21:21:03 +0800</pubDate><description><![CDATA[LLM推理MCQA时存在选择偏差问题（模型依赖选项"A/B/C/D"而非内容本身作答，模型爱选C）

提出名为PriDe（Prior Debiasing）推理时使用无标签（不需要答案）的无监督去偏方法]]></description><category></category></item><item><title><![CDATA[ANSWER, ASSEMBLE, ACE: UNDERSTANDING HOW LMS ANSWER MULTIPLE CHOICE QUESTION 2025ICLR]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/144671430</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/144671430</guid><author>Conger_2002</author><pubDate>Sat, 06 Dec 2025 21:03:38 +0800</pubDate><description><![CDATA[通过机械解释性（Mechanistic Interpretability）的方法，揭示模型如何在内部“克服”符号偏差并完成答案选择的。ANSWER, ASSEMBLE, ACE: UNDERSTANDING HOW LMS ANSWER MULTIPLE CHOICE QUESTION 2025ICLR]]></description><category></category></item><item><title><![CDATA[Github Readme 基本格式语法]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/155618797</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/155618797</guid><author>Conger_2002</author><pubDate>Fri, 05 Dec 2025 23:41:57 +0800</pubDate><description><![CDATA[Github Readme 基本格式语法]]></description><category></category></item><item><title><![CDATA[The SkipSponge Attack: Sponge Weight Poisoning of Deep Neural Networks]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/144754252</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/144754252</guid><author>Conger_2002</author><pubDate>Thu, 26 Dec 2024 22:15:52 +0800</pubDate><description><![CDATA[The SkipSponge Attack: Sponge Weight Poisoning of Deep Neural Networks]]></description><category></category></item><item><title><![CDATA[Stealthy Attack on Large Language Model based Recommendation]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/144672163</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/144672163</guid><author>Conger_2002</author><pubDate>Mon, 23 Dec 2024 16:45:17 +0800</pubDate><description><![CDATA[将LLM引入推荐系统 recommender systems (RS)，存在潜在脆弱性：

攻击者可以在测试阶段改变目标项目的文本内容（标题、描述）来提高目标项目的曝光率，而不需要直接干扰模型的训练过程。攻击很隐蔽的，不会影响整体推荐性能。]]></description><category></category></item><item><title><![CDATA[SMOOTHLLM Defending LLM Against Jailbreaking Attacks (1)]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/144671998</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/144671998</guid><author>Conger_2002</author><pubDate>Mon, 23 Dec 2024 16:36:10 +0800</pubDate><description><![CDATA[越狱攻击：通过**设计输入** 欺骗模型 生成不当内容。]]></description><category></category></item><item><title><![CDATA[SAFEGUARD IS A DOUBLE-EDGED SWORD: DENIALOF-SERVICE ATTACK ON LARGE LANGUAGE MODELS]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/144671885</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/144671885</guid><author>Conger_2002</author><pubDate>Mon, 23 Dec 2024 16:32:36 +0800</pubDate><description><![CDATA[新型拒绝服务（DoS）攻击。与传统的Dos不同，这种攻击利用了LLM安全机制中的假阳性问题。]]></description><category></category></item><item><title><![CDATA[DENIAL-OF-SERVICE POISONING ATTACKS ON LARGE LANGUAGE MODELS]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/144671718</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/144671718</guid><author>Conger_2002</author><pubDate>Mon, 23 Dec 2024 16:31:10 +0800</pubDate><description><![CDATA[基于数据投毒的DoS攻击（P-DoS），通过注入设计用于DoS目的的恶意样本，突破输出长度限制。]]></description><category></category></item><item><title><![CDATA[COERCING LLMS TO DO AND REVEAL (ALMOST) ANYTHING]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/144671496</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/144671496</guid><author>Conger_2002</author><pubDate>Mon, 23 Dec 2024 16:27:02 +0800</pubDate><description><![CDATA[COERCING LLMS TO DO AND REVEAL (ALMOST) ANYTHING

针对llm的对抗性攻击目标不止越狱，攻击目标还有  误导、模型控制、拒绝服务或数据提取。]]></description><category></category></item><item><title><![CDATA[欢迎使用Markdown编辑器]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/144671455</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/144671455</guid><author>Conger_2002</author><pubDate>Mon, 23 Dec 2024 16:23:16 +0800</pubDate><description><![CDATA[.md文件使用超全语法笔记]]></description><category></category></item><item><title><![CDATA[Certifying LLM Safety against Adversarial Prompting]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/144671240</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/144671240</guid><author>Conger_2002</author><pubDate>Mon, 23 Dec 2024 16:21:12 +0800</pubDate><description><![CDATA[Certifying LLM Safety against Adversarial Prompting]]></description><category></category></item><item><title><![CDATA[**Adversarial Demonstration Attacks on Large Language Models**]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/144670837</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/144670837</guid><author>Conger_2002</author><pubDate>Mon, 23 Dec 2024 16:17:30 +0800</pubDate><description><![CDATA[提出**advICL**，操控示例demonstration不改变输入，误导模型。发现对抗性示例增加，icl的鲁棒性降低。

提出**Transferable advICL**，Transferable advICL生成的对抗示例可以攻击未见过的测试输入样例。]]></description><category></category></item><item><title><![CDATA[overleaf 写论文 语法笔记]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/144291966</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/144291966</guid><author>Conger_2002</author><pubDate>Mon, 09 Dec 2024 14:42:54 +0800</pubDate><description><![CDATA[点击绿色按钮“创建新项目”后，可以新建空白项目，可以选择模板直接往模板里添加内容,或者可以上传自己的项目。latex格式的公式可以搜在线latex编辑，（可以图片转公式，但只有三次免费使用机会）注册账户后并登录后进入这个界面，创建的所有历史项目会在这个界面显示，后期可以继续修改。宏包后, 我们就可以用这样的命令切换颜色, 我们一般使用。有些时候图片太大，放在双栏部分图片难以完全且很好的展示出来，的公式，用$ $符号        $宏包并配合特定颜色名字来设定颜色.\textbf{文本}]]></description><category></category></item><item><title><![CDATA[AAAI—24—Main—paper（关于Multi—Modal的全部文章摘要）]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/142884902</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/142884902</guid><author>Conger_2002</author><pubDate>Sat, 12 Oct 2024 19:38:48 +0800</pubDate><description><![CDATA[我们生活在一个由多种模态（Multimodal）信息构成的世界，包括视觉信息、听觉信息、文本信息、嗅觉信息等等，当研究的问题或者数据集包含多种这样的模态信息时我们称之为多模态学习多模态机器学习旨在处理学习（视觉，听觉，语言等）不同模态融合交织的信息。]]></description><category></category></item><item><title><![CDATA[python的常用模块，必能覆盖你的需求]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/141954952</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/141954952</guid><author>Conger_2002</author><pubDate>Sun, 08 Sep 2024 16:31:53 +0800</pubDate><description><![CDATA[Request，Numpy，SciPy，Pillow，SOLAIchemy]]></description><category></category></item><item><title><![CDATA[2024年“研究生科研素养提升”系列公益讲座在线测试]]></title><link>https://blog.youkuaiyun.com/Conger_2002/article/details/141404476</link><guid>https://blog.youkuaiyun.com/Conger_2002/article/details/141404476</guid><author>Conger_2002</author><pubDate>Wed, 21 Aug 2024 22:44:35 +0800</pubDate><description><![CDATA[本测试为2024年“研究生科研素养提升”系列公益讲座课后测试，观看课程总时长大于15小时/B站签到6次，目通过在线测试(60分以上)可获得本次讲座课时证明。参与考试需注意： 1. 在参与考试时，请勿开启浏览器翻译插件，请勿开启VPN或校外访问，否则将影响答题或交卷。 2. 请务必使用桌面浏览器进行考试（桌面浏览器是指谷歌、火狐、edge等电脑端浏览器），请勿使用微信浏览器或使用手机打开本页面，请勿在用手机微信扫描二维码进入手机端网址后，再复制到电脑浏览器打开页面。 如果在考试中遇到问题，请联系知网研学在线客]]></description><category></category></item></channel></rss>