Daily Research Brief 2026-05-13

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

A 26M-parameter model achieved on-device function calling; AI ‘self-replication’ success jumped 13x to 81%; three ministries issued top-level agent policy; DeepSeek V4 broke the million-token context window; Anthropic crossed $900B valuation, eyeing an October IPO.

1. Latest arXiv Papers

1. TBA: Decoupling LLM RL Training — 50x Speedup — https://arxiv.org/abs/2503.18929

2. IntentGrasp: The First Comprehensive ‘Intent Understanding’ Benchmark — https://arxiv.org/abs/2605.06832

3. PRISM: A Three-Stage Multimodal Model Training Framework — https://arxiv.org/abs/2604.28123

4. OPUS: Theory-Guided Pretraining Data Selection for LLMs — https://arxiv.org/abs/2602.05400

5. LLaVA-CKD: Cascaded Knowledge Distillation for Vision-Language Models — https://arxiv.org/abs/2605.10641

6. Continual Factual Knowledge Acquisition in Language Models — https://arxiv.org/abs/2605.10640

7. Confidence-Guided Diffusion Augmentation for Low-Resource Settings — https://arxiv.org/abs/2605.10916

8. Fast Rates for Offline Contextual Bandits with Function Approximation — https://arxiv.org/abs/2605.10639