Daily Research Brief 2026-08-05

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

Two main threads twisted into one today: domestic models pressurize on both ‘capability’ and ‘capital’, while overseas giants are forced to respond with ‘price cuts’. DeepSeek restarts a ¥50B funding round, Moonshot Kimi enters a Pre-IPO round, and Kimi K3, MiniMax H3, ByteDance Seedance 2.5 and DeepSeek-V4-F push capability while the incumbents cut prices.

1. Latest arXiv Papers

1. UrbanAgent: A Tool-Augmented Agent for Cross-System Urban Tasks — https://arxiv.org/abs/2608.03018

2. Control Under Compression: Reliability Frontiers for Tool-Using Agents — https://arxiv.org/abs/2608.01056

3. Verifiable Memory: Learning Unified Memory Management for Long-Horizon LLM Agents — https://arxiv.org/abs/2608.03137

4. A Game Theory for Foundation Models: Inferring Similarity via Embedded Equilibrium — https://arxiv.org/abs/2608.03958

5. EmbodiedVAE: Disentangled Video VAE for Efficient and Controllable Embodied Manipulation — https://arxiv.org/abs/2608.02990

6. RoboReact: Agentic Skill Distillation from Generated Egocentric Videos — https://arxiv.org/abs/2608.03387

7. Ego2Robot: Scalable Robot Data Synthesis from Egocentric Human Data — https://arxiv.org/abs/2608.02580

8. IBA-Bench: An Interactive Benchmark for Assessing AI Agents — https://arxiv.org/abs/2608.02109