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.

Impact：Data and privacy leaks Impact：Availability and reliability Layer：Model layer Layer：Training and data layer Clear all

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

Evaluation and benchmarksarXiv 2610.12758Oct 9
DistressBench: Evaluating multi-dimensional crisis support in large language models
Introduces DistressBench, evaluating models' multidimensional support in crisis conversations
arXiv
2610.12758
Published
Oct 9
Layer
Model layer
Setting
Models and APIs
Evaluated models
Muse Spark 1.2, Muse Spark 1.3, Muse Glimmer 30B and others20
AbstractThe model is more restrained and acts less.
DistressBench uses SSH dialogues written and revised by clinical personnel, and evaluates how much necessary care general chat models deliver according to weighted atomic criteria, rather than whether they refuse. When users disclose suicidal ideation or self-harm to a chat model, the authors consider the expected behavior to be recognizing the crisis, calibrating empathy, de-escalating, and referring to human help. Refusal-based evaluations cannot detect failing to act after recognizing the crisis.
Paper analysisRead analysis
DefensesarXiv 2610.12602Oct 9
Self-Preservation in Language Models
Proposes Task Bank self-replay to mitigate held-out forgetting in continual instruction tuning
arXiv
2610.12602
Published
Oct 9
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen3-0.6B, Qwen3-1.7B, Qwen3-4B and others4
- DefensesModel hardening
AbstractThe authors claim self-replay can reduce forgetting outside the list, and that on the 4B model new task information helps with target selection, while smaller models do not.
The authors study how language models, when learning new tasks, can use their own knowledge for broader self-preservation: given an incomplete list and forgetting that extends beyond the list, can the model generate replay and help choose targets.
Paper analysisRead analysis
DefensesarXiv 2610.10655Oct 8
Nullify：面向 LLM 遗忘的零空间激活引导免训练方法
提出零空间激活引导 Nullify，免训练擦除指定记忆并保持效用
arXiv
2610.10655
Published
Oct 8
Layer
Model layer
Setting
Models and APIs
Evaluated models
Llama-2-7B-chat-hf, Qwen3-8B
- DefensesInference-time intervention
- AttacksExtraction and theft
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.10276Oct 7
PatchBench 提出激活补丁局部评测协议，揭示越狱修复的附带损伤
提出 PatchBench-Local 协议，评测激活补丁修复越狱的局部附带损伤
arXiv
2610.10276
Published
Oct 7
Layer
Model layer
Setting
Models and APIs
Evaluated models
Qwen2.5-3B-Instruct, Qwen2.5-14B-Instruct, gemma-3-4b-it and others8
- DefensesInference-time intervention
- AttacksJailbreaks
Abstract论文提出 PatchBench 与局部评估协议，揭示了激活补丁在全局能力无损下仍存局部良性退化。
Paper analysisRead analysis
AttacksarXiv 2610.09819Oct 7
FAB 攻击：给 HuBERT/WavLM 植入后门，背景警笛声即可让下游语音任务失灵
提出 FAB 攻击，向声学基础模型植入可由环境声触发的后门
arXiv
2610.09819
Published
Oct 7
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
HuBERT-base, WavLM-base
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentAudio
Abstract论文提出了针对声学基模的 FAB 后门攻击，在弱威胁模型下实现了任务无关、物理可触发的后门植入。
Paper analysisRead analysis
OtherarXiv 2610.08917Oct 7
CARE：为视觉-语言-动作模型推理加速提供认证式选择
提出 CARE，离线认证 VLA 加速配置的回合级失败风险并选出最快合格候选
arXiv
2610.08917
Published
Oct 7
Layer
Model layer
Setting
Embodied systems and robotics
Evaluated models
OpenVLA-OFT, π0.5, π0 and others5
AbstractCARE 以配对回合认证加速诱发失败，在预算内选最快候选，不足则保留参考。
CARE 在部署前从加速配置中选出最快、且加速诱发失败风险能被有限样本认证为低于用户预算的候选。证据不足就保留全计算参考策略。
Paper analysisRead analysis
Analysis and theoryarXiv 2610.08577Oct 6
学习方式如何影响遗忘：记忆与泛化谱系下的机器遗忘研究
分析记忆与泛化如何决定机器遗忘的 retain 损伤
arXiv
2610.08577
Published
Oct 6
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
one-layer ReLU transformer, one-layer transformer (BMA, width doubled), Hubble-8B-100B perturbed and others6
Abstract作者称共享规则占比越高，遗忘带来的 retain damage 越大。
作者考察学习方式如何左右 unlearning 的 retain damage：训练越依赖跨样本共享的解法，在匹配的遗忘程度上，retain 上的损失越大。
Paper analysisRead analysis
DefensesarXiv 2610.06636Oct 6
差分隐私混合公共数据集提升隐私学习效果
提出 DP-MixMin，私有学习公开数据混合并做差分隐私微调
arXiv
2610.06636
Published
Oct 6
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
ResNet-18, GPT-2 Medium
- AttacksExtraction and theft
AbstractDP-MixMin 私下学公开混合再 DP 微调，胸片 macro AUC 与邮件困惑度优于均匀混合。
DP-MixMin 是一条端到端的差分隐私训练流程：先私下决定多个公开数据集的预训练混合，再预训练并在敏感数据上 DP 微调。作者称这是首个这样的流程。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.06171Oct 5
ControlPed 生成逼真行人危险动作场景，七款端到端驾驶模型 HDScore 从 88.8 降至 47.4
提出 ControlPed 与 HazardPed，生成可控行人危险场景并评测端到端驾驶
arXiv
2610.06171
Published
Oct 5
Layer
Model layer
Setting
Embodied systems and robotics
Evaluated models
DiffusionDrive, SSR, VAD and others7
- ComponentVideo
Paper analysisRead analysis
AttacksarXiv 2610.06083Oct 5
研究者提出面向深度强化学习的轨迹级可迁移黑盒对抗攻击
提出轨迹级可迁移对抗攻击，降低深度强化学习策略的期望回报
arXiv
2610.06083
Published
Oct 5
Layer
Model layer
Setting
Embodied systems and robotics
Evaluated models
DQN, Double DQN (DDQN)
- TechniqueGradient and representation optimization
Abstract轨迹级后退视界攻击在 CartPole 上比逐步 FGSM 类和 Sign 噪声更压低回报。
这项工作研究 DRL 上的迁移式黑盒观测攻击：受害参数不可访问时，在白盒代理上造扰动再迁移。
Paper analysisRead analysis
OtherarXiv 2610.05831Oct 5
长时程轨迹监督如何影响终端智能体训练的可靠性与成本
提出选择性长时程精炼，筛选轨迹以提升终端智能体训练可靠性
arXiv
2610.05831
Published
Oct 5
Layer
Training and data layer
Setting
coding agent
Evaluated models
Qwen/Qwen3-8B
Abstract中间长度的监督更均衡。
这项工作研究终端智能体模仿学习里，教师轨迹应保留多少 token。
Paper analysisRead analysis
AttacksarXiv 2610.05601Oct 5
Tracer 框架可从遗忘后的扩散模型中识别被擦除概念
提出 TRACER，从权重差识别扩散模型被擦除概念
arXiv
2610.05601
Published
Oct 5
Layer
Model layer
Setting
Models and APIs
Attacker
White-box
Evaluated models
Stable Diffusion v1.5, Stable Diffusion 3, FLUX.1 and others5
- AttacksExtraction and theft
- ComponentImage generation
AbstractTRACER 用权重足迹识别未披露擦除概念并估计个数，作者称数秒内完成且准确率高于对照。
TRACER 针对的是遗忘扩散模型上的目标识别：提供者从公开基座 W 用已发表方法擦掉未知集合 S⋆，只发布 W′。攻击者还有假定包含这些概念的词表，但不知擦了谁、擦了几个、用了什么算法。
Paper analysisRead analysis
AttacksarXiv 2610.05453Oct 5
研究者提出 Privacy-Leaking Watermarks，可在本地多模态模型中泄露对话隐私
提出 PLW，用微调把先前对话敏感触发编码进后续生成图像
arXiv
2610.05453
Published
Oct 5
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
OmniGen2, BAGEL-7B
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentFine-tuning and open weights
AbstractPLW 把先前对话线索写成后续图像中的隐水印，检出随触发、模型和间隔变化。
Paper analysisRead analysis
Analysis and theoryarXiv 2610.05076Oct 4
自生成反馈会破坏测试时训练：长时程适应的因果分解
因果分解自生成反馈如何破坏测试时训练的长时程适应
arXiv
2610.05076
Published
Oct 4
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
125M TTT-E2E, 760M TTT-E2E, 3B TTT-E2E and others5
Abstract自写闭环损害真实文本预测。
持久测试时更新若来自模型自己生成的文本，会同时改变模型和后续训练文本。作者在 PG-19 长流上分解这条闭环，并用独立文本决定是否提交更新。
Paper analysisRead analysis
DefensesarXiv 2610.04998Oct 4
EmoRSS：缓解大语言模型由情绪引发的过度拒答
提出 EmoRSS，推理时在拒绝子空间反向抵消情绪激活偏移以缓解过度拒答
arXiv
2610.04998
Published
Oct 4
Layer
Model layer
Setting
Models and APIs
Evaluated models
Llama-3.1-8B-it, Qwen3-8B
- DefensesInference-time intervention
- RiskInappropriate refusal
Paper analysisRead analysis
OtherarXiv 2610.04792Oct 4
多模态一定更好吗？缺失模态鲁棒性的几何视角与 Geodesic Unlearning 方法
提出 Geodesic Unlearning，沿测地线编辑权重以提升缺模态鲁棒性
arXiv
2610.04792
Published
Oct 4
Layer
Model layer
Setting
Models and APIs
Evaluated models
RoBERTa, ResNet50, Qwen2.5-3B and others6
- ComponentVision
AbstractGU 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。
Paper analysisRead analysis
DefensesarXiv 2610.04616Oct 3
PerturBot 用扰动训练打破视觉-语言-动作模型的模态捷径
提出 Perturbot 扰动训练，打破 VLA 的模态捷径
arXiv
2610.04616
Published
Oct 3
Layer
Training and data layer
Setting
Embodied systems and robotics
Evaluated models
π0.5
- DefensesModel hardening
- ComponentVision
AbstractPerturbot 改训练数据以削弱三类模态捷径，GroundFscore 用来看成功是否来自任务证据。
Perturbot 是一套不改推理的 VLA 训练期数据干预，用来削弱模态捷径，并用离线分数区分“任务变好”和“靠捷径变好”。
Paper analysisRead analysis
DefensesarXiv 2610.04426Oct 3
UnAct：无需梯度的定向激活干预实现类别遗忘
提出 UnAct，仅以前向激活干预从分类器中遗忘指定类别
arXiv
2610.04426
Published
Oct 3
Layer
Model layer
Setting
Models and APIs
Evaluated models
ResNet-18, ViT-B/16
- DefensesModel hardening
- AttacksExtraction and theft
AbstractUnAct 用前向激活做类别遗忘。
UnAct 是一种事后类别遗忘：删除请求可以只有少量该类图像，且没有标签和保留数据。它不用梯度，只对遗忘图像做前向。编辑对象是分类器所读的后期单元。
Paper analysisRead analysis
DefensesarXiv 2610.04299Oct 3
R-Quest：用问题有效性与新颖性反馈维持推理模型自我进化
提出 R-Quest，用有效性与新颖性反馈维持推理模型自我进化
arXiv
2610.04299
Published
Oct 3
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen3-4B-Base, OctoThinker-3B-Hybrid-Base
- DefensesModel hardening
- RiskReward hacking
AbstractR-Quest 用有效性与新颖性反馈维持自进化。
R-Quest 处理的是推理模型自进化中的题目质量：questioner 与 solver 用 GRPO 交替从自生成题目学习时，多轮之后平均成绩会下降。
Paper analysisRead analysis
AttacksarXiv 2610.04128Oct 3
研究显示拆分学习中的梯度可提升客户端文本重建率
衡量分裂学习中梯度相对激活对客户文本重建的额外贡献
arXiv
2610.04128
Published
Oct 3
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
GPT-2 (openai-community/gpt2, 124M), Qwen3-0.6B
- AttacksExtraction and theft
Abstract激活已恢复大部分 token。
这篇工作测量 split learning 分裂点上的文本泄漏：激活单独能恢复多少，训练回传的梯度再增加多少，以及按 token 还是按整篇计分会不会改写结论。
Paper analysisRead analysis

AI Safety Daily

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