AI News Digest 2026-08-18

直近2日間のAI関連ニュースをまとめました。

Today's Features

Digest

• DeepSeek V4 Pro at 207 tok/s with the full 1M context, no quantization [HN DeepSeek/Grok/OpenCode / importance 92 / dev 60]
• Grok 4.6 vs GPT-5.6 Sol 2026: Which AI Model Wins? - Memeburn [Google News Grok/xAI / importance 92 / dev 75]
• Grok 4.6 matches GPT-5.6 Sol on benchmarks while undercutting its own $300 plan - Startup Fortune [Google News Grok/xAI / importance 92 / dev 50]
• Claude Fable 5 highとGPT-5.6 Sol medium、ハルシネーションが少ないのはどちらか [Zenn LLM / importance 92 / dev 0]
• DeepSeek AI Releases DeepSeek Harness in Developer Preview: An MIT-Licensed Agent Harness Where Everything is a Plugin - MarkTechPost [Google News DeepSeek / importance 85 / dev 85]
• AWS Open-Sources Dogwood, Extending Cedar to Govern Sequences of Agent Tool Calls [InfoQ (AI/ML/Data Eng) / importance 85 / dev 90]
• OpenAI signs record Ohio data center lease with Nvidia backing up to $105 billion [The Decoder / importance 85 / dev 60]
• HELIX: Model-Harness Co-evolution for Recursive Self-Improvement [arXiv cs.AI / importance 85 / dev 95]
• Engineering Reliable Coding Agents: Evaluating and Operating the System Around the Model [arXiv cs.AI / importance 85 / dev 90]
• TENET: One Step Toward Test-Driven Development for Repository-Level Code Generation [arXiv cs.AI / importance 85 / dev 90]
• VISOR: Agentic Visual Retrieval-Augmented Generation via Iterative Search and Over-horizon Reasoning [arXiv cs.AI / importance 85 / dev 90]
• PhoneWorld: Scaling Phone-Use Agent Environments [arXiv cs.AI / importance 85 / dev 90]
• From Code Review to Code Critique: Intent, Drift, and Spotlight for AI-Generated Diffs at Scale [arXiv cs.AI / importance 85 / dev 90]
• Agentao: A Governed Local-First Runtime for Tool-Using LLM Agents [arXiv cs.AI / importance 80 / dev 90]
• Ontology-Grounded Project Memory for Coding Agents [arXiv cs.AI / importance 80 / dev 90]
• Demystifying Agent Skills: Why They Work-Until They Don't [arXiv cs.AI / importance 80 / dev 90]
• SKILLER: Language-Level Reinforcement Learning for Reusable Skill Extraction in Small Language Models [arXiv cs.AI / importance 80 / dev 85]
• Adaptive Stopping for Multi-Turn LLM Reasoning [arXiv cs.AI / importance 80 / dev 85]
• TIDE: Proactive Multi-Problem Discovery via Template-Guided Iteration [arXiv cs.AI / importance 80 / dev 85]
• Kalypso: Relational LLM Serving [arXiv cs.AI / importance 80 / dev 90]
• Stripe is reportedly acquiring AI startup OpenRouter for more than $7 billion [The Decoder / importance 78 / dev 50]
• Llama.cpp v0.1.0 [Hacker News / importance 75 / dev 90]
• Nvidia investing $1.5B in SoftBank data center developer behind OpenAI project [TechCrunch (AI) / importance 75 / dev 50]
• AI Used to Verify Toughest Mathematics Proof Yet [IEEE Spectrum (AI) / importance 75 / dev 70]
• Evaluating Agentic Learning Harness Capabilities Without Labels via the Scaling Hypothesis [arXiv cs.AI / importance 75 / dev 85]
• MemoryLake on MemoryArena: A Matched Study of Agent Memory Backends [arXiv cs.AI / importance 75 / dev 90]
• Agent-Orchestration in Autonomous Chip Design [arXiv cs.AI / importance 75 / dev 90]
• Mandato: Protocol-Level Enforcement of Digitally Signed Mandates on AI Agent Actions with Cryptographically Chained Audit Trails [arXiv cs.AI / importance 75 / dev 90]
• A Graph-Based Reinforcement Learning Framework for Structured Drift Diagnosis and Recovery in Autonomous LLM Agents [arXiv cs.AI / importance 75 / dev 85]
• From BERT to Frontier Agents: Eight Years of Language-Model Progress, the Collapse of the Capability-Cost Curve, and the Rise of Task-Targeted Models [arXiv cs.AI / importance 75 / dev 80]
• Engineering Signals of Human-AI Collaboration in the Agentic Coding Era: A Longitudinal Analysis of 33,228 Pull Requests from vLLM and SGLang with Implications for Biomedical AI Agents and Bioinformatics Pipeline Developmen [arXiv cs.AI / importance 75 / dev 80]
• Agentic Transaction: Towards ACID-Compliant Agent Systems [arXiv cs.AI / importance 75 / dev 85]
• GRASP: Gated Regression-Aware Skill Proposer for Self-Improving LLM Agents [arXiv cs.AI / importance 75 / dev 80]
• SkillSight: Calibrating Generic Content Bias for Skill Retrieval [arXiv cs.AI / importance 75 / dev 80]
• PhyAI: Real-Time Physical AI at the Edge, Scalable Rollouts in the Cloud [arXiv cs.AI / importance 75 / dev 85]
• Why we need an AI-resilient society [arXiv cs.AI / importance 75 / dev 55]
• Edge Case Detection in Automated Driving: Methods, Challenges, and Future Directions [arXiv cs.AI / importance 75 / dev 75]
• INFORM-CT: INtegrating LLMs and VLMs FOR Incidental Findings Management in Abdominal CT [arXiv cs.AI / importance 75 / dev 80]
• AtomBridge: Agentic VLA Inference Plugin for Long-Horizon Tasks in Scientific Experiments [arXiv cs.AI / importance 75 / dev 85]
• Early Stopping for Large Reasoning Models via Confidence Dynamics [arXiv cs.AI / importance 75 / dev 85]
• Fine-grained Claim-level RAG Benchmark for Law [arXiv cs.AI / importance 75 / dev 85]
• RL-Index: Reinforcement Learning for Retrieval Index Reasoning [arXiv cs.AI / importance 75 / dev 85]
• ShortOPD: Recovering Pruned LLMs with Short-to-Long On-Policy Distillation [arXiv cs.AI / importance 75 / dev 85]
• RepBench: Compiling Benchmarks into Capability Representations for Large Language Models [arXiv cs.AI / importance 75 / dev 85]
• Cloud-ScPO: Hidden-State Geometry for Semi-Supervised Preference Optimization in LLM Reasoning [arXiv cs.AI / importance 75 / dev 85]
• DeepSeek Releases V4 Pro With Higher Benchmarks, Open-Source Tooling, and Upcoming Price Increases - gHacks [Google News DeepSeek / importance 70 / dev 60]
• Groq raises $350M to fuel its pivot from AI chips to neocloud [TechCrunch (AI) / importance 70 / dev 55]
• Article: Agentic Fitness Functions: Extending Evolutionary Architecture Beyond Deterministic Rules [InfoQ (AI/ML/Data Eng) / importance 70 / dev 85]
• Modular Cognitive Architecture Emerges in Large Language Models [arXiv cs.AI / importance 70 / dev 85]
• Measuring Cross-Task Behavioral Consistency in Language Model Agents [arXiv cs.AI / importance 70 / dev 85]
• Second Thought: Reasoning in Parallel as LLM Agents Act and Observe [arXiv cs.AI / importance 70 / dev 85]
• AI Research Preference Models [arXiv cs.AI / importance 70 / dev 85]
• ScienceFlow: A long-horizon agent for ML research, scientific discovery and beyond [arXiv cs.AI / importance 70 / dev 80]
• A Four-Axis Trustworthiness Benchmark for LLM-as-Judge in Principle-Based Regulation [arXiv cs.AI / importance 70 / dev 75]
• DeaMoE: Efficient MoE Structure for Fast Small-Batch Decoding [arXiv cs.AI / importance 70 / dev 85]
• Ensuring Safe Physical AI in Urban Mobility via Hazard-Informed Synthesized Envelopes [arXiv cs.AI / importance 70 / dev 70]
• Ads in AI Chatbots? An Analysis of How Large Language Models Navigate Conflicts of Interest [arXiv cs.AI / importance 70 / dev 60]
• NEURON: A Neuro-symbolic System for Grounded Clinical Explainability [arXiv cs.AI / importance 70 / dev 70]
• Making AI-Generated Feedback Matter: A Large-Scale Study of Feedback Workflows and Student Enactment [arXiv cs.AI / importance 70 / dev 65]
• OTIS: Learning High-Quality Time Series Features With Tiny Encoders [arXiv cs.AI / importance 70 / dev 80]
• The Fools are Certain; the Wise are Doubtful: Exploring LLM Confidence in Code Completion [arXiv cs.AI / importance 70 / dev 80]
• A Systematic Comparison of Training Objectives for Out-of-Distribution Detection in Image Classification [arXiv cs.AI / importance 70 / dev 80]
• SomaliBench Eval: Measuring English-to-Somali Refusal Gaps in Open-Weight Language Models [arXiv cs.AI / importance 70 / dev 75]
• Breaking Chains with Trees: Model-Parallel Deep Learning with $\mathcal{O}(\log N)$ Time Complexity [arXiv cs.AI / importance 70 / dev 85]
• Learning Dexterous Manipulation Using Contact Wrench Guidance From Human Demonstration [arXiv cs.AI / importance 70 / dev 80]
• TypeProbe: Recovering Type Representations from Hidden States of Pre-trained Code Models [arXiv cs.AI / importance 70 / dev 85]
• Teffic-Audio: Tell Fact from Fiction [arXiv cs.AI / importance 70 / dev 80]
• ATLAS: Discovering Agent Strategies through LLM-Guided Abstraction and Automata Learning [arXiv cs.LG / importance 70 / dev 85]
• AI Agent Guardrails for Real-World Actions [Dev.to LLM / importance 70 / dev 85]
• OpenAIのエージェント群に想定外の「チーム」が生まれた [Zenn LLM / importance 70 / dev 80]