1
Tracing Agentic Failure from the Flow of Success
从成功流程中追溯智能体失败根源，揭示AI自主决策的脆弱性
arXiv:2607.12747v1 Announce Type: cross Abstract: Failure attribution for LLM-based agentic systems, i.e., identifying which steps in a failure trajec…
2
Self-Study Reconsidered: The Hidden Fragility of Learning from Self-Generated QA
揭示语言模型自我生成QA训练的隐藏脆弱性，一篇值得关注的AI研究论文。
arXiv:2606.32002v1 Announce Type: new Abstract: Language models are increasingly taught from synthetic question--answer (QA) supervision: a model gene…
3
Safety Is Not Universal: The Selective Safety Trap in LLM Alignment
揭秘LLM安全评估的致命缺陷：看似全面防护实则暗藏针对特定群体的盲区，戳破“选择性安全陷阱”系统性风险
arXiv:2601.04389v3 Announce Type: replace-cross Abstract: Current safety evaluations of large language models (LLMs) create a dangerous illusion of un…
4
How Fragile Are Training-Free AI-Generated Image Detectors? A Controlled Audit of Score Direction, Preprocessing, and Compression
揭秘无训练AI图像检测器的脆弱性，系统评估分数方向、预处理与压缩的三重影响。
arXiv:2606.20488v1 Announce Type: new Abstract: Training-free detectors of AI-generated images promise generator-agnostic deployment without classifie…
5
The Autonomy Tax: Defense Training Breaks LLM Agents
论文揭示防御训练会让LLM智能体付出“自主性税”，性能与安全如何平衡？
arXiv:2603.19423v2 Announce Type: replace-cross Abstract: Large language model (LLM) agents increasingly rely on external tools (file operations, API …
6
PromptMN: Pseudo Prompting Language
借助伪提示语言（Pseudo Prompting Language），系统化解决自然语言提示词中角色、目标与约束松散导致的交互脆弱性，提升生成式AI与智能体对齐效率。
arXiv:2606.17164v1 Announce Type: cross Abstract: Prompting has become the primary interface between humans and generative AI, yet many natural langua…
7
Cognitive Debt: AI as Intellectual Leverage and the Dynamics of Systemic Fragility
这篇arXiv论文提出“认知债务”概念，剖析AI作为智力杠杆如何引发系统性脆弱性，值得技术决策者深思。
arXiv:2606.15078v1 Announce Type: new Abstract: We develop a formal theory of cognitive debt: the stock of unverified reasoning obligations that accum…
8
Acquisition state behaves as a structured, measurable variable governing lung-nodule AI: kernel-driven measurement instability and noise-driven detection fragility, invisible to DICOM metadata
揭示医学影像AI在肺部结节检测中受采集状态影响的内在不稳定性，为AI治理提供定量依据
arXiv:2606.12824v1 Announce Type: cross Abstract: AI governance for medical imaging is formalizing: the 2026 ACR-SIIM Practice Parameter recommends lo…
9
When Probing Accuracy Saturates, Fragility Resolves: A Complementary Metric for LLM Pre-Training Analysis
当标准探测准确率饱和时，引入“脆弱性”度量作为互补指标，为LLM预训练分析提供新视角。
arXiv:2606.11375v1 Announce Type: cross Abstract: Standard linear probing declares a property "encoded" when a classifier on hidden states achieves hi…
10
When Autoregressive Consistency Hurts Safety Alignment
自回归一致性让大模型安全对齐脆弱不堪：微调只能重塑输出开头几个token，后续轨迹难以纠正。
arXiv:2606.04168v1 Announce Type: new Abstract: Safety alignment in large language models (LLMs) is fragile in part because it is often shallow: fine-…
11
Linear Ensembles Wash Away Watermarks: On the Fragility of Distributional Perturbations in LLMs
揭示LLM水印在多重模型访问下的致命缺陷：独立扰动轻松被线性集成抹除，对AI安全与版权保护提出新挑战。
arXiv:2605.30501v1 Announce Type: new Abstract: Watermarking embeds statistical signatures in AI-generated text for detection and attribution. We reve…
12
Short-Term Gain, Long-Term Fragility: AI Labor Substitution and the Erosion of Sustainable Capability
这篇论文揭示AI替代人力的短期效率提升，实则埋下长期能力丧失的隐患，是反思技术与社会关系的必读研究。
arXiv:2605.27399v1 Announce Type: cross Abstract: What looks like acceleration can be a quiet transfer of burden from the present to the future. Attem…
13
The Fragility of Chain-of-Thought Monitoring Across Typologically Diverse Languages
多语言环境下，大模型思维链监控因语言类型差异暴露脆弱性，AI安全研究新视角。
arXiv:2605.27901v1 Announce Type: cross Abstract: Chain-of-thought (CoT) monitoring has been proposed as a promising safety mechanism for detecting mi…
14
Understanding the Challenges in Iterative Generative Optimization with LLMs
揭示LLM迭代优化中的脆弱性：仅9%的智能体能成功自改进，挑战何在？
arXiv:2603.23994v2 Announce Type: replace-cross Abstract: Generative optimization uses large language models (LLMs) to iteratively improve artifacts (…
15
Unveiling the Fragility of Vision-Language Models: Multi-Modal Adversarial Synergy via Texture-Constrained Perturbations and Cross-Modal Optimization
顶会AAAI 2026论文，揭示视觉语言模型在多模态对抗攻击下的脆弱性，提出纹理约束与跨模态优化的协同方法。
arXiv:2605.26501v1 Announce Type: cross Abstract: Large Vision-Language Models (LVLMs) have transformed multi-modal understanding, excelling in tasks …
16
Constraint Decay: The Fragility of LLM Agents in Back End Code Generation
arXiv新研究揭示LLM代理在后端代码生成中面临"约束衰减"问题，系统脆弱性令人警醒。
Article URL: https://arxiv.org/abs/2605.06445 Comments URL: https://news.ycombinator.com/item?id=48256912 Points: 4 # Comments: 0
17
On Robustness and Chain-of-Thought Consistency of RL-Finetuned VLMs
首份系统研究RL微调VLM的鲁棒性与思维链一致性，揭示模型脆弱性根源
arXiv:2602.12506v3 Announce Type: replace Abstract: Reinforcement learning (RL) finetuning has become a key technique for enhancing large language mod…
18
Alignment Dynamics in LLM Fine-Tuning
揭秘LLM微调中对齐为何脆弱：从参数动态到输出分布的统一视角
arXiv:2605.18309v1 Announce Type: new Abstract: Although Large Language Models (LLMs) achieve strong alignment through supervised fine-tuning and rein…
19
Adversarial Fragility and Language Vulnerability in Clinical AI: A Systematic Audit of Diagnostic Collapse Under Imperceptible Perturbations and Cross-Lingual Drift in Low-Resource Healthcare Settings
临床AI系统在细微扰动和多语言场景下存在诊断崩溃风险，这篇系统性审计揭开了安全漏洞。
arXiv:2605.16993v1 Announce Type: cross Abstract: Current clinical artificial intelligence (AI) systems are evaluated almost exclusively on clean, sta…
20
On the Fragility of Data Attribution When Learning Is Distributed
分布式学习中数据归因的脆弱性：单个参与者可操纵归因值大幅膨胀，挑战定价与审计可信度。
arXiv:2605.15520v1 Announce Type: cross Abstract: Data attribution has become an important component of pricing, auditing, and governance in machine l…