Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
All papers Found 63 papers Another 305 papers do not have category tags yet and are not included in filtered results.
“
About this collection Collection in progress We are building a large AI safety paper collection to help researchers find work related to their interests. Papers are collected automatically andfiltered only for relevance to AI safety , without judging quality, novelty or reading value. The default “Annotated first” shows papers with tags or published paper analyses first. Each group is ordered by original publication date, newest first, with undated papers last. “Newest first” uses publication dates only. This order is not a quality ranking.
One-line summaries, tags and paper analyses are prepared by AI from the papers. Language models may have systematic biases and produce hallucinations.We plan to carry out human annotation and review to reduce errors in this content.
Propose LWDL, a measure sparse dictionary learning method under linearized optimal transport
Published Oct 9 Abstract LWDL learns sparse measurement dictionaries in flat embedding spaces.LWDL embeds probability measures into a Hilbert space, then performs sparse dictionary learning.
Propose studying world models, using language models to predict intervention gains
Published Oct 9
Evaluated models Claude Opus 5 , Grok-4.6 , GPT-6 Astra and others13 Abstract The authors claim experimental records can be reused inside and outside the environment and improve experiment selection under fixed budgets.The authors use language models as research world models, predicting the median intervention gain before execution, to compare candidate experiments under a limited budget. The gap to address is that research agents propose faster than they execute. The same type of change can have different effects across reference models, recipes, and budgets, so experimental knowledge that can continue to be used outside the original environment is needed.
比对捐赠对话史与众包提交,测量 AI 代答的集中度与获批
Published Oct 7 给出线性 robust bandits 特例的多项式时间学习者并证明小推广 NP-hard
Published Oct 7 Abstract 欧氏球仿射特例可多项式时间达到Õ(√T)。这是一篇关于不可实现 bandit 学习之计算边界的理论工作:线性 robust bandits 的一个欧氏球仿射特例有多项式时间、Õ(√T)遗憾的学习者,再做小推广就会落到复杂性坍塌。
综述部署时参数内记忆的注入位点与获取时机
Published Oct 7 Abstract 两轴加三个算子整理部署时插入的参数记忆,并标出干扰与安全等开放问题。部署时插入前向的参数记忆被单独划界,并按位点与时机归类。
系统综述青年 AI 安全风险与对策的证据阶段
Published Oct 6 Abstract 作者称多数风险-对策关联停在提议,并主张把目标扩向安全的青年发展。
主张多模态安全评测应衡量可控性,并用 SAE 建立四轴画像
Published Oct 5
Evaluated models LlavaGuard-v1.2-7B-OV , Qwen3.5-9B Abstract 四轴画像把读出与选择性控制分开。作者主张多模态安全评测在行为分类之外,单独报告可控性画像,而不是把检测准确当成控制证据。
把文本水印证据写成带错误率的三态可审计结论
Published Oct 5 Abstract 把水印合规从普遍检测改成带范围的证据曲线、三态报告和再校准,模拟支持平方修正。指定配置下的文本水印被做成可审计测量:记录证据如何衰减,并把检测收成三态,而不是判定人类作者身份。
用伤害风险模型评估英国 AI 生成虚假信息的伤害潜能
Published Oct 5 提出任务感知的跨域蒸馏与剪枝方法以压缩音频深度伪造检测器
Published Oct 4
Evaluated models XLSR-AASIST 提出 Geodesic Unlearning,沿测地线编辑权重以提升缺模态鲁棒性
Published Oct 4
Evaluated models RoBERTa , ResNet50 , Qwen2.5-3B and others6 Abstract GU 用测地线子空间编辑应对部署时缺模态下降,并尽量保住全模态精度。这项工作讨论多模态模型在部署时整段缺一个模态的性能下降,并把校正写成训练后的子空间编辑。
提出 DecSteer,用共享低秩残差算子把审议收成直接决策
Published Oct 3
Evaluated models Qwen3.5-4B , Llama-3.2-3B-Instruct , Qwen3.5-0.8B Abstract DecSteer 用小参数 BC 算子把长推理收成短决策,增益随基座余量变化并可组合。DecSteer 是加在冻结语言模型残差流上的 System-1 决策算子,用行为克隆代替强化学习来调用模型里已有的技能。
提出一阶转向,把可组合权重适配译成可相加激活转向
Published Oct 3
Evaluated models Qwen2.5-VL-3B , Gemma-4-E2B , Mistral-7B-Instruct-v0.3 提出 Patch-to-Prune,旁路 VLM 部分层的视觉投影计算
Published Oct 2
Evaluated models Qwen2.5-VL-3B-Instruct , Qwen2.5-VL-7B-Instruct , LLaVA-v1.6-Vicuna-7B-HF and others4 Abstract P2P 用深度上的激活修补旁路视觉投影,以可调容忍度交换准确率与计算量。P2P 是面向 VLM 推理的训练无关计算旁路:不删除视觉 token,而在选定 decoder 层用固定激活代理代替视觉投影。
提出 LS-AR,用潜向量经 FiLM 调节自回归解码以保持宏观目标
Published Oct 2
Evaluated models LS-AR (Static) , LS-AR (Dynamic) Abstract 用潜通道把宏观目标从 token 流拆出,并在合成检索与规划上比较静态、动态两种调节。LS-AR 给因果解码器加一条不占窗口的连续目标通道,用来处理长程生成里宏观指令被挤出上下文或被注意力稀释的问题。
提出 HEST,用熵训练的双曲探针只在犹豫 token 引导隐状态
Published Oct 2
Evaluated models Qwen2.5-Math-1.5B-Instruct , Qwen2.5-Math-7B-Instruct , Llama-3.1-8B-Instruct Abstract 只改高熵 token 的双曲转向,多数设置高于 greedy,全 token 固定向量则下降。HEST 是不更新语言模型权重的推理时转向:模型自己的 next-token 熵既选出少数犹豫 token,也训练给出方向的双曲探针。要处理的情况是,数学解答里多数 token 已由上下文决定,分叉集中在高熵 token,而常见激活加法仍对每个 token 加同一欧氏向量。长推理模型上按关键词和步骤边界转向的做法,又很少适用于 instruction-tuned 模型。
提出 TESS 与点值匹配损失,训练可迁移的数据选择网络
Published Oct 2
Evaluated models Llama-3-8B-Instruct , Llama-3.1-8B-Instruct , Qwen2.5-7B-Instruct and others4 Abstract TESS 用损失差伪标签和 PVM 训练可迁移选择网络,并在安全选择与指令微调上报告跨数据迁移。TESS 是一个用于 LLM 训练数据选择的元学习框架:用验证集引导下的损失差做样本伪标签,训练一个能给未见样本打分的选择网络。
证明随机谕示下谕示辅助计算无法零知识验证,签名谕示下则可
Published Oct 2 Abstract 随机谕示排除一般零知识监督。一般谕示辅助计算在随机谕示中做不到 witness hiding,因而也做不到零知识监督。给谕示答案加上数字签名后,抗碰撞哈希足以同时得到隐私和简洁验证。
评述全球 AI 治理,认为问题是行动者权力与激励分歧
Published Oct 1 Abstract 书评认为形成中的架构由多类行动者塑造,但没有共同计划。
提出 AI 评估沙盒配置器,支撑监管沙盒多方技术评估与报告
Published Oct 1
Evaluated models LLM-backed conversational assistant Abstract 框架把每次技术测试配成沙盒。AI Assessment Sandbox Configurator 是面向欧盟 AIRS 中结构化技术测试的开源配置框架,用来按一次参与组装评估环境。