Daily Research Brief 2026-06-18

📊 Token usage: estimated from retrieval and writing scale.

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

Editor’s Note

SSE released audit guidelines for the STAR Market’s fifth listing standard for AI models — opening an IPO fast lane for unprofitable model companies; DeepSeek closed a ¥51B Series A at a ¥400B valuation with Tencent, JD and CATL participating. On the paper side: MemTrace long-term memory, SDS-LoRA gradient anisotropy, and agentic discovery of cardiac digital twins.

1. Latest arXiv Papers

1. Equal-Position Trajectories: Revealing the Internal Mechanisms of LLM Arithmetic Errors — https://arxiv.org/abs/2606.03645

2. MemTrace: An Evaluation Framework for LLM Long-Term Memory — https://arxiv.org/abs/2606.17328

3. Quantifying LLM Logical-Reasoning Consistency via Structural Uncertainty — https://arxiv.org/abs/2606.17312

4. Can LLMs Autonomously Discover the Mathematical Concept of Zero? — https://arxiv.org/abs/2606.17289

5. DELM: A Decentralized Multi-Agent Collaboration Framework for Language Models — https://arxiv.org/abs/2606.10662

6. SDS-LoRA: Overcoming Gradient Anisotropy in LoRA Fine-Tuning — https://arxiv.org/abs/2606.16454

7. End-to-End Battery Health Prediction via Temporal Contrastive Representation Learning — https://arxiv.org/abs/2606.16434

8. Cardiac Electrophysiology Digital Twins via Autonomous Hybrid-Structure Discovery — https://arxiv.org/abs/2606.18154