Daily Research Brief 2026-07-21

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

Two threads worth practitioners’ attention today. First, ‘open-weight models enter a dense payoff period’ — DeepSeek V4 officially GA’d open-source (1.6T MoE, fully MIT-licensed), Qwen3.8 went open (2.4T), and China’s Meteorological Administration open-sourced a hundred-billion-parameter weather model tied to global public early warning; combined with Kimi K3 weights dropping 7/27, the open camp is advancing on parameter scale, domain specialization and usability simultaneously — the ‘open = catching up’ narrative has been substantively overturned this week. Second, agents are descending from the chat box into infrastructure primitives.

1. Latest arXiv Papers

1. Reward-Driven LLM Agent Workflows: Synthesizing POMDP Routing and Self-Correction for Autonomous Decision-Making — https://arxiv.org/abs/2607.17038

2. Dynamic Agent Skills: A Lifecycle Survey and Taxonomy of Evolving Skill Libraries — https://arxiv.org/abs/2607.10113

3. AgentAbstain: Do LLM Agents Know When Not to Act? — https://arxiv.org/abs/2607.10059

4. Isolation as a First-Class Principle for LLM-Agent System Safety: Concepts, Taxonomy, Challenges and Future Directions — https://arxiv.org/abs/2607.12406

5. Critic Experience Bank: Self-Evolving Step-Level Confidence Estimation for LLM Agents — https://arxiv.org/abs/2607.12397

6. PM-Bench: Evaluating Prospective Memory in LLM Agents — https://arxiv.org/abs/2607.12385

7. On-Device Deep Research at 4B: Exposure Bounds Faithfulness, Retrieval Bounds Coverage — https://arxiv.org/abs/2607.12257

8. A Threshold Exceedance Framework for CBRN Uplift Evaluation in Frontier Language Models — https://arxiv.org/abs/2607.12200