92 reports from 2025–2026, organized by major model families and concrete engineering questions · 92 papers · 中文 ↗
Selection rule: technical detail is still required, but company and model-family prominence now matters. Well-known adjacent versions are kept when comparing them reveals how architecture, data, post-training, modalities, or serving changed; a famous name alone is still not enough, and the false-attribution Llama 4 arXiv entry remains excluded.
0 / 92 read
How to read: record the company, model version, constraint, main design choice, and evidence boundary. Compare adjacent versions before deciding that a newer report supersedes an older one, and treat reported benchmark or production gains as conditional on the stated data, hardware, harness, and workload.