Daily Research Brief 2026-05-20

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

The paper thread today: agentic discovery for test-time compute (LLMs Improving LLMs), normalizing trajectory models, conformal path reasoning for trustworthy KGQA, Mixture-of-Experts pretraining (UniPool/EMO), and why global LLM leaderboards mislead.

1. Latest arXiv Papers

1. LLMs Improving LLMs: Agentic Discovery for Test-Time Compute — https://arxiv.org/abs/2605.08083

2. Normalizing Trajectory Models — https://arxiv.org/abs/2605.08078

3. Conformal Path Reasoning: Trustworthy KGQA via Path-Based Conformal Prediction — https://arxiv.org/abs/2605.08077

4. GRAPHLCP: Structure-Aware Localized Conformal Prediction — https://arxiv.org/abs/2605.08074

5. STARFlow2: Bridging Language Models and Normalizing Flows — https://arxiv.org/abs/2605.08021

6. UniPool: A Globally Shared Expert Pool for Mixture-of-Experts — https://arxiv.org/abs/2605.06665

7. EMO: Pretraining Mixture of Experts for Emergent Modularity — https://arxiv.org/abs/2605.06663

8. Crafting Reversible SFT Behaviors in Large Language Models — https://arxiv.org/abs/2605.06632

9. Why Global LLM Leaderboards Are Misleading — https://arxiv.org/abs/2605.06656

10. MASPO: Joint Prompt Optimization for LLM-Based Multi-Agent Systems — https://arxiv.org/abs/2605.06641