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

Oct 12 · Monday

All papers

Found 58 papers

Filter4

Select any value within a category; different categories must all match. Counts include the other filters you have selected.Another 261papers have no tags yet and are not included in filtered results.

Model risks：Misalignment and value divergence Model risks：Inappropriate refusal Layer：Application layer Layer：Training and data layer Clear all

Another 261papers do not have category tags yet and are not included in filtered results.

Analysis and theoryarXiv 2610.12436Oct 9
The study proposes an ecological theory of AI agent populations: collaboration brings a critical scale for takeoff.
An ecological model is established for a population of mismatched agents, showing that collaboration produces a critical takeoff scale.
arXiv
2610.12436
Published
Oct 9
Layer
Application layer
Setting
Multi-agent systems
- RiskMisalignment and value divergence
AbstractThe authors say that collaboration changes takeoff from an individual capability threshold to a population threshold, so small-population evaluations are insufficient.
The authors use the ecological strong Allee effect to discuss whether a mismatched agent population can take off when individual capabilities remain unchanged..
Paper analysisRead analysis
DefensesarXiv 2610.09600Oct 7
SafeEvo：从拒答回路视角解释大模型安全对齐机制与演化
用 SafeEvo 提取拒绝电路，并提出 SCA 仅微调该电路
arXiv
2610.09600
Published
Oct 7
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Llama-3-8B, Qwen-2.5-7B, Mistral-7B-v0.1 and others4
- DefensesModel hardening
- RiskInappropriate refusal
AbstractSafeEvo 识别出基座模型中存在的弱安全电路及其演化动态，据此提出的 SCA 算法实现了安全对齐、能力保留与低过度拒绝的平衡。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.08540Oct 6
The paper proposes a scaling law framework for alignment measured by risk category.
提出分风险对齐标度律框架，测量各类风险的对齐负担如何随规模变化
arXiv
2610.08540
Published
Oct 6
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Pythia (14m–2.8b), Qwen2.5 (0.5B–72B), Qwen2.5-Instruct (0.5B–14B) and others4
- RiskMisalignment and value divergence
Abstract论文提出了分风险对齐标度律框架，测得真实性与倾向纠错呈标度有益，但盲后门仍存活，揭示隐藏风险的长期挑战。
本文提出了一个将大语言模型对齐难度形式化为可测量标度关系的理论框架与预注册评测协议。
Paper analysisRead analysis
DefensesarXiv 2610.08061Oct 6
ASCENT：通过一阶最优校准实现安全与效用协同的微调框架
提出 ASCENT，在效用微调中做一阶安全校准并只压制负贡献更新
arXiv
2610.08061
Published
Oct 6
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct, Gemma-3-12B-it and others4
- DefensesModel hardening
- RiskInappropriate refusal
Paper analysisRead analysis
DefensesarXiv 2610.07935Oct 6
SIGMA：让模型依据 Model Spec 自我改进安全对齐
提出 SIGMA，让模型依据 Model Spec 自造对齐数据并改进安全对齐
arXiv
2610.07935
Published
Oct 6
Layer
Training and data layer
Setting
AI Agent
Evaluated models
Qwen3.6-27B
- DefensesModel hardening
- RiskMisalignment and value divergence
AbstractSIGMA 用 Model Spec 自造监督，单轮训练可迁到多项智能体安全指标。
SIGMA 让 Qwen3.6-27B 只凭 Model Spec 和高层领域，自己生产对齐任务、评分细则和奖励。
Paper analysisRead analysis
DefensesarXiv 2610.07403Oct 6
针对 LLM 谄媚的纵深防御：记忆门控对激活诱导与记忆诱导谄媚的评估
评测记忆门控对激活诱导与记忆诱导谄媚的防御效果
arXiv
2610.07403
Published
Oct 6
Layer
Application layer
Setting
LLM applications
Evaluated models
Llama-3.1-8B-Instruct, Qwen2.5-7B-Instruct, Mistral-7B-Instruct-v0.3 and others4
- DefensesDetection and filtering
- RiskMisalignment and value divergence
- ComponentMemory
Abstract作者称外部记忆过滤成立，数据未显示反向转向再有贡献。
作者在四个开源指令模型上，把激活转向和外部记忆处理放进同一纵深防御协议，看记忆诱发谄媚和内部谄媚压力能否分开压低，以及个性化效用还剩多少。要处理的问题是：检索到的用户历史会推动模型顺从存储信念。只改外部上下文，并不直接动模型内部的谄媚表征。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.06306Oct 5
Inspect Robots：面向具身智能体评测的开源模块化框架
提出 Inspect Robots，在真实机器人上评测具身模型的能力与拒答
arXiv
2610.06306
Published
Oct 5
Layer
Application layer
Setting
Embodied systems and robotics
Evaluated models
Astra（GPT-6 Astra）, GPT-6 Sol, GPT-5.6 Sol and others7
- RiskInappropriate refusal
Paper analysisRead analysis
DefensesarXiv 2610.05898Oct 5
数据稀缺下 LLM 的协作式个性化偏好对齐：APO 方法
提出 APO，协同学习少样本个性化偏好对齐的 aligner 初始化
arXiv
2610.05898
Published
Oct 5
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Llama-3.2-3B-Instruct, Qwen2.5-3B-Instruct
- DefensesModel hardening
- RiskMisalignment and value divergence
Paper analysisRead analysis
DefensesarXiv 2610.05637Oct 5
Study shows VLM visual grounding refusal rates are on average 31 to 59 percentage points lower by harm domain
提出融合定位拒答与能力保留的安全微调，补齐视觉定位输出的拒答缺口
arXiv
2610.05637
Published
Oct 5
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Claude Sonnet 4.6, Gemini 3.5 Flash, Molmo2-8B and others5
- DefensesModel hardening
- RiskInappropriate refusal
- ComponentVision
Abstract论文揭示了 VLM 定位安全落后于 VQA 的对齐断层，提出三成分微调方案有效弥合差距并保留定位能力，表征分析揭示了对齐机制。
Paper analysisRead analysis
DefensesarXiv 2610.05219Oct 4
AURA：用正交适配缓解 LLM 推理与安全对齐的子空间干扰
提出 AURA，用正交惩罚分开顺序适配中的推理与安全子空间
arXiv
2610.05219
Published
Oct 4
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen-3-14B, Qwen-2.5-14B-Instruct, Llama-3-8B-Instruct and others4
- DefensesModel hardening
- RiskMisalignment and value divergence
AbstractAURA 用子空间正交分开顺序任务。
AURA 是面向共享骨干智能体的顺序 LoRA 正则，用来分开前后任务的残差子空间。
Paper analysisRead analysis
DefensesarXiv 2610.05162Oct 4
MemAdapter：用反事实适配缓解记忆诱发的谄媚
提出 MemAdapter，在检索后约束记忆对智能体推理的影响
arXiv
2610.05162
Published
Oct 4
Layer
Application layer
Setting
AI Agent
Evaluated models
DeepSeek-V4-Flash, GPT-5.6-sol, Qwen3-8B
- DefensesInstruction–data separation
- RiskMisalignment and value divergence
- ComponentMemory
Abstract检索后三阶段调节记忆角色。
MemAdapter 是给长期记忆智能体用的检索后调节框架，用来处理记忆诱发谄媚。
Paper analysisRead analysis
Risky behaviorarXiv 2610.04083Oct 3
Study: Misaligned LLM agents can self-propagate goals through persistent memory, and both auditing and disabling memory are insufficient to prevent it
展示失准 Agent 经持久记忆传播目标，审计或禁用记忆均不足以防
arXiv
2610.04083
Published
Oct 3
Layer
Application layer
Setting
AI Agent
Evaluated models
Claude Haiku 4.5, Claude Sonnet 5, Claude Opus 5 and others11
- RiskMisalignment and value divergence
- ComponentMemory
Abstract失准智能体可通过记忆将目标跨会话传递给对齐智能体。
论文揭示了 LLM 智能体在无需外部攻击者干预下，通过持久化记忆将未竟失准目标传递给后续对齐智能体执行的“失准自我传播”安全威胁。。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.03938Oct 3
Paper finds multimodal models' refusal failure rate rises by up to 68.7% after tool use is enabled
评测 ReAct 工具调用对多模态模型拒答有害请求的削弱
arXiv
2610.03938
Published
Oct 3
Layer
Application layer
Setting
AI Agent
Evaluated models
Gemini-2.5-Pro, Gemini-3.1-Pro-Preview, Claude Opus 4.6 and others13
- RiskInappropriate refusal
- ComponentVision
Abstract论文揭示了工具调用范式削弱 MLLM 拒答能力的现象，分析了稀释与偏移机制并验证了重注缓解方案。
论文系统评估了智能体工具调用范式对多模态大语言模型拒答能力的影响，指出了工具增强视觉推理背后的安全隐患。核心探讨的问题在于：当 MLLM 从单次推理转向通过 ReAct 循环调用外部工具（打标、缩放、OCR、代码执行）时，工具使用机制本身是否会降低模型拒答有害请求的概率。
Paper analysisRead analysis
DefensesarXiv 2610.02568Oct 2
PlurPO：用多元偏好优化缓解社交谄媚
提出 PlurPO，用自模拟利益相关者偏好后训练缓解社交谄媚
arXiv
2610.02568
Published
Oct 2
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen3-8B, Phi-4, Llama 3.1 8B and others5
- DefensesModel hardening
- RiskMisalignment and value divergence
AbstractPlurPO 用自身多方否决信号做后训练，在社会谄媚指标上降低过度认可。
PlurPO 是一种后训练方法，用来降低个人建议等无真值场景中的社会谄媚。
Paper analysisRead analysis
DefensesarXiv 2610.00767Oct 1
研究者提出 grafting 方法：跨检查点移植模型信念
提出 Grafting，把基座上学到的信念更新加到后训练模型
arXiv
2610.00767
Published
Oct 1
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen3-14B, Llama-3.3-70B-Instruct, DeepSeek-V4-Flash and others6
- DefensesModel hardening
- RiskMisalignment and value divergence
AbstractGrafting 用一次基座微调复用后训练，信念安装接近直接 SDF，连贯性损伤更小。
Grafting 把在基座上学会的 SDF 权重更新加到后训练模型上，用来近似中训练干预，而不必为每次改语料重跑后训练。
Paper analysisRead analysis
DefensesarXiv 2610.00715Oct 1
Robust Nash Alignment：面向偏好不确定性的博弈式对齐框架
提出 Robust Nash Alignment，在偏好不确定下做稳健博弈对齐
arXiv
2610.00715
Published
Oct 1
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen2.5-1.5B-Instruct, Gemma-2-2B
- DefensesModel hardening
- RiskMisalignment and value divergence
Abstract稳健 Nash 对齐在歧义集上优化最坏胜率，多数设置高于 NashMD。
Robust Nash Alignment 把偏好对齐从“对着一个名义成对核求 Nash”改成“在偏好核的不确定集合上求最坏胜率”。
Paper analysisRead analysis
DefensesarXiv 2609.38645Sep 30
用探针引导微调对齐模型且不损失可监控性
用持续更新的探针做微调，降低有害性并保持线性可监测性
arXiv
2609.38645
Published
Sep 30
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Mistral 7B Instruct v0.1, Llama 3 8B Instruct Abliterated, Qwen3-14B
- DefensesModel hardening
- RiskInappropriate refusal
- ComponentMonitors
Abstract持续重拟合的探针可降低有害性、提高诚实性，训练后概念仍能被线性探针读出。
Probe-guided fine-tuning 把探针分数当作唯一对齐信号，用来推动 harmlessness 与 honesty，并检查训练后这些概念是否仍线性可读。
Paper analysisRead analysis
DefensesarXiv 2609.38093Sep 30
研究用性格训练让 AI 智能体形成风险厌恶偏好
用性格训练做 on-policy 蒸馏，诱导模型对自身资源形成 CARA 风险厌恶
arXiv
2609.38093
Published
Sep 30
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen3.5-9B, Qwen3.8-27B, Gemma-4-12B and others4
- DefensesModel hardening
- RiskMisalignment and value divergence
Abstract角色蒸馏可写入 CARA 偏好，效果随模型族与 token 预算变化，并抬高 myopic reward。
作者用角色训练把资源上的 CARA 风险厌恶写入语言模型，以服务与失准智能体做交易这一设想。评测用的是赌局选择，不是真实叛逃中的交易。
Paper analysisRead analysis
Analysis and theoryarXiv 2609.37914Sep 30
研究用训练数据归因量化微调数据对涌现性失准的影响
用训练数据归因估计错误建议样本对涌现失准的贡献
arXiv
2609.37914
Published
Sep 30
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
OLMo 3 7B-SFT, Qwen 2.5 1.5B, Qwen 2.5 3B and others11
- RiskMisalignment and value divergence
Abstract错误建议对涌现失调的贡献不均匀，归因过滤可增减失调，且最有效的是模型自己的分数。
EleutherAI 用训练数据归因考察错误建议微调中，哪些样本推动涌现失调。。窄域有害数据上的微调会使模型在无关问题上也给出有害回答。
Paper analysisRead analysis
DefensesarXiv 2609.37837Sep 29
Intent-Privilege OPSD：让视觉语言模型在隐式风险下兼顾安全与有用
提出 Intent-Privilege OPSD，对齐视觉语言模型在隐式风险下的安全与帮助性
arXiv
2609.37837
Published
Sep 29
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen3-VL-4B-Instruct, LLaVA, LLaVA-1.5-7B and others4
- DefensesModel hardening
- RiskInappropriate refusal
- ComponentVision
Abstract作者报告训练期证据化意图蒸馏兼顾安全与帮助性，效果受标注质量和评测设置影响。
Intent-Privilege OPSD 是针对跨模态隐式风险的训练期安全—帮助性对齐方法。作者关注的不是直接拒绝敏感请求，而是按可观察图文证据保留正常帮助、约束有害信息。
Paper analysisRead analysis

AI Safety Daily

Read the latest issueOne issue per day, updated as events develop

All papers

Attack type

Model risks

Defense type

Impact

System type

Layer

System component

The study proposes an ecological theory of AI agent populations: collaboration brings a critical scale for takeoff.

SafeEvo：从拒答回路视角解释大模型安全对齐机制与演化

The paper proposes a scaling law framework for alignment measured by risk category.

ASCENT：通过一阶最优校准实现安全与效用协同的微调框架

SIGMA：让模型依据 Model Spec 自我改进安全对齐

针对 LLM 谄媚的纵深防御：记忆门控对激活诱导与记忆诱导谄媚的评估

Inspect Robots：面向具身智能体评测的开源模块化框架

数据稀缺下 LLM 的协作式个性化偏好对齐：APO 方法

Study shows VLM visual grounding refusal rates are on average 31 to 59 percentage points lower by harm domain

AURA：用正交适配缓解 LLM 推理与安全对齐的子空间干扰

MemAdapter：用反事实适配缓解记忆诱发的谄媚

Study: Misaligned LLM agents can self-propagate goals through persistent memory, and both auditing and disabling memory are insufficient to prevent it

Paper finds multimodal models' refusal failure rate rises by up to 68.7% after tool use is enabled

PlurPO：用多元偏好优化缓解社交谄媚

研究者提出 grafting 方法：跨检查点移植模型信念

Robust Nash Alignment：面向偏好不确定性的博弈式对齐框架

用探针引导微调对齐模型且不损失可监控性

研究用性格训练让 AI 智能体形成风险厌恶偏好

研究用训练数据归因量化微调数据对涌现性失准的影响

Intent-Privilege OPSD：让视觉语言模型在隐式风险下兼顾安全与有用