Daily Research Brief 2026-06-28

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

The US declared API inference access an export-controlled activity, tightening frontier-model regulation; AICE 2026 opened in Guangzhou covering the full AI value chain. On the paper side: JetFlow’s 3x speculative decoding, GauntletBench multimodal generalization, the co-failure ceiling proof across 67 frontier models, and Unlimited-OCR (40-page single-pass OCR).

1. Latest arXiv Papers

1. The Overlooked Dividend in LLM Post-Training: Agent Step-Level Evaluation Without Reward Models — https://arxiv.org/abs/2606.27891

2. Why Multi-Step Tool-Use RL Collapses: Supervision-Signal Repair for Catastrophic Failure — https://arxiv.org/abs/2606.28147

3. JetFlow: Breaking the Scaling Ceiling of Speculative Decoding — 3x Long-Text Inference — https://arxiv.org/abs/2606.27563

4. GauntletBench: A Multimodal Generalization Benchmark That Takes Agents Out of Their Comfort Zone — https://arxiv.org/abs/2606.28329

5. The Co-Failure Ceiling of 67 Frontier Models: A Proof of the Accuracy Upper Bound of Model Ensembles — https://arxiv.org/abs/2606.27942

6. DSpark: Semi-Autoregressive Inference Acceleration — 85% Faster in High-Concurrency Scenarios — https://arxiv.org/abs/2606.28671

7. Unlimited-OCR: A 3B-Parameter MoE Long-Document OCR Model Handling 40 Pages at Once — https://github.com/baidu/Unlimited-OCR

8. A Theoretical Framework for Multi-Agent Collaboration Efficiency — 217% Efficiency Gain