The article presents LoGeR, a long‑context geometric reconstruction framework that combines test‑time‑training memory and sliding‑window attention to achieve minute‑scale, fully‑feedforward 3D reconstruction with superior accuracy on benchmarks such as KITTI and VBR.
The article examines Cursor’s Composer 2 model, detailing its self‑summarization reinforcement‑learning workflow, the limitations of traditional compression methods, token‑efficient results on the CursorBench benchmark, and a challenging Terminal‑Bench case study that demonstrates dramatically reduced token usage while improving performance.
The article details Qwen3's three‑phase pretraining pipeline, long‑context extensions, a cold‑start long‑chain‑of‑thought dataset, reinforcement‑learning fine‑tuning with custom rewards, and a two‑stage distillation process that yields versatile, thought‑controlled language models.
The article presents a practical method for evaluating 1‑million‑token LLMs by inserting structured medal data into a classic Chinese novel, provides a full Python script for the test, shares results on GLM‑4‑long, and discusses training techniques and open‑source resources for long‑context models.
