Daily Research Brief 2026-07-20

📊 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 today. First, ‘open-weight models reach the substantive stage of matching closed source’ — Thinking Machines’ Inkling (975B), PrismML squeezing 27B into an iPhone, DeepSeek V4 set for 7/24, plus the FLI safety index finding ‘open vs closed gap narrowed to 4-7 months’ — open source is no longer a cheap alternative but a capability competitor, and enterprise model selection must include open weights as a default candidate. Second, ‘agents descend from the chat box into infrastructure primitives’.

1. Latest arXiv Papers

1. RxBrain: Embodied Cognition Foundation Model with Joint Language-Visual Reasoning and Imagination — https://arxiv.org/abs/2607.14187

2. Think at 5 Hz, Act at 20 Hz: Asynchronous Fast-Slow Vision-Language-Action Inference for Closed-Loop Driving — https://arxiv.org/abs/2607.15621

3. S-squared-VLA: Decoupling Semantic and Spatial Streams in Vision-Language-Action Models for Autonomous Driving — https://arxiv.org/abs/2607.13926

4. Scaling Behavior Foundation Model for Humanoid Robots — https://arxiv.org/abs/2607.15163

5. FVAttn: Adaptive Sparse Attention with Runtime Load Balancing for Video Generation — https://arxiv.org/abs/2607.16190

6. MotionForesight: Re-purposing Video Models for Future 3D Scene-Flow Prediction — https://arxiv.org/abs/2607.16192

7. ReBind: Multi-Reference Video Editing via Structured Instructions with Explicit Reference Relationships — https://arxiv.org/abs/2607.14681

8. StructAgent: Harness Long-horizon Digital Agents with Unified Causal Structure — https://arxiv.org/abs/2607.11388