Daily Research Brief 2026-08-12

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

This week’s main thread is the descent from ‘model capability race’ to ‘agent infrastructure build-out’: Meta ships 30B Muse Glimmer as an open-weight model runnable on consumer GPUs, Cloudflare/Tencent Cloud open-source agent runtime foundations like ‘sandbox computers’ and ’team memory’ — small and mid teams now get the same runtime base.

1. Latest arXiv Papers

1. LDR: Extrapolative Video World Models via Latent Dynamics Reasoning — https://arxiv.org/abs/2608.09926

2. SCOUT: Self-Checking and Recoverable Tool-Reasoning Agent for Ultra-Long First-Person Videos — https://arxiv.org/abs/2608.07959

3. AutoPrune: LLM-Automated Design of Visual Token Pruning Strategies — https://arxiv.org/abs/2608.07193

4. AnyTrack: Unified Visual Object Tracking Framework Across Modalities — https://arxiv.org/abs/2608.06773

5. AgentPatch: Training-Free Coarse-to-Fine Repair with Merged Agent Multimodal LLMs — https://arxiv.org/abs/2608.06699

6. Unified Agent: Stateful Interaction Management Across Devices and Time — https://arxiv.org/abs/2608.05729

7. OneDayAgent: Long-Horizon Orchestration Framework for Autonomous Agents — https://arxiv.org/abs/2608.05013

8. Video-DeepResearch: Toward Next-Generation Multimodal Deep Research Agents — https://arxiv.org/abs/2608.03979