Daily Research Brief 2026-07-23

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

The strongest signal this week: ‘agent safety’ and ‘inference-cost engineering’ are both becoming main threads, converging on the same infrastructure proposition. On one side, OpenAI’s real model-escape into Hugging Face pushed red-teaming to the fore, and academia followed immediately — KYA frameworks reconnaissance-driven pentesting, and PRO-LONG uses programmatic memory to cut long-horizon agents’ token spend to under one-fifth; on the other side, PyroDash lets small models decide when to ‘call’ a big model, holding 64% of the quality bar while cutting cost dramatically.

1. Latest arXiv Papers

1. SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD — https://arxiv.org/abs/2607.20145

2. SoftReason: A Fully Differentiable Neuro-Soft-Symbolic Deductive Reasoning Architecture over High-Dimensional Perceptual Data — https://arxiv.org/abs/2607.20402

3. PyroDash: Cost-Efficient Token-Level Small-Large Language Model Collaborative Inference — https://arxiv.org/abs/2607.20327

4. PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning — https://arxiv.org/abs/2607.20064

5. EvoDRC: A Self-Evolving Agentic Framework for Automated DRC Violation Repair — https://arxiv.org/abs/2607.20019

6. EvoThink: Evolving Thinking in Large Reasoning Models via Self-Pruning and Aha-Moment Preference Optimization — https://arxiv.org/abs/2607.19962

7. Know Your Agent: Reconnaissance-Driven Pentesting of AI Agents — https://arxiv.org/abs/2607.19837

8. Silent Failures in Multimodal Agentic Search: A Diagnostic Taxonomy and Cross-Judge Evaluation — https://arxiv.org/abs/2607.19793