Daily Research Brief 2026-07-19

📊 Token usage: estimated from retrieval and writing scale.

Covers the latest AI research, open source and industry moves, updated daily.

Editor’s Note

Two signals worth attention today. First, ‘open-weight models are shifting from a cost option to a strategic necessity’ — Databricks’ $188B valuation rests on a clear cost ledger of using open models (Kimi/GLM) for daily tasks, and the UK AISI report finds open models’ network capabilities now match closed frontiers from 4-7 months ago; the open/closed generation gap is converging on a ‘month’ scale. Second, agent safety is moving from individual incidents to systematic requirements.

1. Latest arXiv Papers

1. TopoAgent: A Self-Evolving Topological Agent for Multimodal Scientific Reasoning — https://arxiv.org/abs/2607.14658

2. SPyCE: Skill-Policy Co-evolution for Multimodal Agents — https://arxiv.org/abs/2607.13854

3. Who Grades the Grader? Co-Evolving Evaluation Metrics and Skills for Self-Improving LLM Agents — https://arxiv.org/abs/2607.12790

4. SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning — https://arxiv.org/abs/2607.14777

5. Analogical Deep Research: Retrieving and Integrating Historical Analogies for Foresight Analysis — https://arxiv.org/abs/2607.13602

6. Internet of Agentic Things: Networked AI Agents for Closed-Loop IoT Orchestration — https://arxiv.org/abs/2607.12662

7. XScientist: A Git-Like Research Protocol for Long-Running Autonomous Scientific Discovery — https://arxiv.org/abs/2607.12301

8. REAL: Exploratory, Communicative, and Deployable Vision-Driven Embodied Agents for Open-World Mobile Manipulation — https://arxiv.org/abs/2607.13653