慣性聚合 関心のあるブログ、ニュース、テクノロジーを効率的に追跡
原文を読む 慣性聚合で開く

おすすめ購読元

The GitHub Blog
The GitHub Blog
T
ThreatConnect
C
Check Point Blog
T
The Exploit Database - CXSecurity.com
U
Unit 42
云风的 BLOG
云风的 BLOG
freeCodeCamp Programming Tutorials: Python, JavaScript, Git & More
T
Tenable Blog
博客园 - 叶小钗
D
Docker
T
Threatpost
WordPress大学
WordPress大学
腾讯CDC
I
Intezer
T
Tailwind CSS Blog
Engineering at Meta
Engineering at Meta
D
Darknet – Hacking Tools, Hacker News & Cyber Security
Hugging Face - Blog
Hugging Face - Blog
cs.CV updates on arXiv.org
cs.CV updates on arXiv.org
The Register - Security
The Register - Security
Stack Overflow Blog
Stack Overflow Blog
PCI Perspectives
PCI Perspectives
S
Security Archives - TechRepublic
Simon Willison's Weblog
Simon Willison's Weblog
A
Arctic Wolf
MongoDB | Blog
MongoDB | Blog
小众软件
小众软件
Hacker News: Ask HN
Hacker News: Ask HN
O
OpenAI News
博客园 - 【当耐特】
L
LINUX DO - 最新话题
C
Comments on: Blog
S
Securelist
月光博客
月光博客
S
Secure Thoughts
Security Latest
Security Latest
MyScale Blog
MyScale Blog
NISL@THU
NISL@THU
F
Full Disclosure
M
Microsoft Research Blog - Microsoft Research
T
True Tiger Recordings
SecWiki News
SecWiki News
aimingoo的专栏
aimingoo的专栏
Threat Intelligence Blog | Flashpoint
Threat Intelligence Blog | Flashpoint
L
LINUX DO - 热门话题
cs.AI updates on arXiv.org
cs.AI updates on arXiv.org
AWS News Blog
AWS News Blog
Hacker News - Newest:
Hacker News - Newest: "LLM"
L
Lohrmann on Cybersecurity
H
Help Net Security

cs updates on arXiv.org

HydraPrompt: An Adaptive and Asymmetric Framework of Vision-Language Models for Synthetic Image Detection Geometry-Aware Representation Denoising for Robust Multi-view 3D Reconstruction 3D Gaussian Map with Open-Set Semantic Grouping for Vision-Language Navigation On the Push-Based Asynchronous Federated Learning: A Bias-Correction Aggregation Approach OmniInteract: Benchmarking Real-World Streaming Interaction for Real-Time Omnimodal Assistants CNNs, Transformers, Hybrid, and Vision Language Models for Skin Cancer Detection VesselSim: learning 3D blood vessel segmentation without expert annotations Erased but Exploitable: Black-box Embedding-Aware Prompting Against Unlearned Text-to-Image Diffusion Models VisualNeedle: Benchmarking Active Visual Search in Information-Dense Scenes DuoGesture: Neuro-Inspired and Biomechanically Informed Dual-Stream Co-Speech Gesture Generation RadarSim: Simulating Single-Chip Radar via Multimodal Neural Fields RoMo: A Large-Scale, Richly Organized Dataset and Semantic Taxonomy for Human Motion Generation The Rescue Effect: Spatio-Semantic Early Exit Bypasses Quantization Collapse in CLIP When Rule Violations Are Rare: Chimera Training for Logical Anomaly Detection Detail Consistent Stage-Wise Distillation for Efficient 3D MRI Segmentation Sparse-LiDAR Prompting of Monocular Geometry Foundations: An Empirical Study Toward Long-Range Driving Depth AirCast-SR: A Foundation Model for Kilometer-Scale Atmospheric Super-Resolution via Latent Consistency Diffusion Personalized Generative Models for Contextual Debiasing Cross-scale Aligned Supervision for Training GANs Joint Instance Segmentation and Geometric Attribute Regression for Roof Structures in Aerial Imagery Dimensional Distribution Emotion State: Leveraging Valence and Arousal as a Common Embedding Space for Visual Emotion Analysis TSFMAudit: Data Contamination Auditing in Forecasting Time Series Foundation Models Clinically-Grounded Counterfactual Reasoning for Medical Video Diagnosis Triadic Dynamics Aware Diffusion Posterior Sampling for Inverse Problems: Optimizing Guidance and Stochasticity Schedules Comparative Study of Vision-Based Metric Measurement for Large-Scale Planar Scenes LongAV-Compass: Towards Unified Evaluation of Minute-Scale Audio-Visual Generation Across T2AV, I2AV, and V2AV The Constraint Tax: Measuring Validity-Correctness Tradeoffs in Structured Outputs for Small Language Models SilIF: Silhouette-Augmented Isolation Forest for Unsupervised Transaction Fraud Detection Multi-Modal Building Inspection via Perceiver IO Fusion of Satellite and Street-Level Imagery E$^3$C: Video Generation with 3D Environmental Memory and Ego-Exo Human Pose Control Share More, Search Less: Collaborative Parallel Thinking for Efficient Test-Time Scaling Unveiling the Fragility of Vision-Language Models: Multi-Modal Adversarial Synergy via Texture-Constrained Perturbations and Cross-Modal Optimization Sleep-stage efficient classification using a lightweight self-supervised model Underwater360: Reconstructing Underwater Scenes from Panoramic Images with Omnidirectional Gaussian Splatting Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective Sentinel: Embodied Cooperative Spatial Reasoning and Planning LongCat-Video-Avatar 1.5 Technical Report GEM: Geometric Entropy Mixing for Optimal LLM Data Curation Uncertainty-Aware Gaussian Map for Vision-Language Navigation Frequency-Guided Fusion For RGB-Thermal Semantic Segmentation BioFact-MoE: Biologically Factorized Mixture of Experts for Vision-Language Prognostic Modeling in Hepatocellular Carcinoma A multifractal-based masked auto-encoder: an application to medical images Benchmarking Convolutional, Transformer, Hybrid, and Vision Language Models for Multi Disease Retinal Screening Unified Panoramic Geometry Estimation via Multi-View Foundation Models Not All Modalities Are Equal: Instruction-Aware Gating for Multimodal Videos OmniGF: A Dual-Branch Vision-Language Framework for Unified Gaze Following Zero-Shot Object Re-Identification in Egocentric Kitchen Videos via Multi-Stage SAM3 Feature Fusion Evi-Steer: Learning to Steer Biomedical Vision-Language Models through Efficient and Generalizable Evidential Tuning Planning Neural Dynamics with Lie Group Embedding through Supervised Projective Manifold Learning AnchorDiff: Training-Free Concept Grounding for MM-DiTs via Anchor-Based Graph Propagation
なぜLLMsは構造化知識で妄想するのか:線形化表現上の推論に関する機械的分析
Shanghao Li, · 2026-05-27 · via cs updates on arXiv.org

PDFを表示 HTML(実験的)

要約:多くの推論タスクにおいて、大規模言語モデル(LLM)はグラフや表などの構造化された外部知識に依存しており、これらは通常、順序的なトークン表現に線形化されます。しかし、十分な知識が利用可能であっても、LLMは依然として幻覚的な出力を生成し、そのような失敗の背後にあるメカニズムは依然としてよく理解されていません。私たちはこれらのメカニズムを調査し、幻覚は体系的な内部ダイナミクスによって引き起こされ、ランダムなノイズではなくということを発見しました。まず、注意は短絡的で似た構造的なヒントに対して不釣り合いに集中し、全体の文脈に均等に分布するのではなく。次に、フィードフォワード表現は提供された知識を定着させることができず、モデルがパラメトリックメモリに戻る原因となります。さらに、私たちの結果は、幻覚は一貫してフィードフォワードレイヤー内の意味の定着の失敗と関連しており、注意の割り当てはよりタスクに依存した変異性を示すことを示しています。最後に、私たちはこれらのメカニズムパターンが単跳びグラフを超えて、マルチ跳びおよび表形式の設定に一般化し、構造化された知識フォーマットを横断する効果的な幻覚検知を可能にすることを示しました。
コメント: ACL 2026 論文集に掲載予定
分野: 計算と言語 (cs.CL); 機械学習 (cs.AI)
報告 番号: ACL 2026
引用: arXiv:2605.26362 [cs.CL]
  (または arXiv:2605.26362v1 [cs.CL] このバージョン用)
  https://doi.org/10.48550/arXiv.2605.26362

DataCiteを通じてarXivから発行されたDOI(登録待ち)

提出履歴

From: Shanghao Li [メールを表示]
[v1] 月曜日、2026年5月25日 22:08:59 UTC (763 KB)