News Around the World — 2026-10-01

AI

• arXiv — Scaling-law study trained 800 models on varying AI-to-human token ratios and finds synthetic web text helps data-starved pretraining at first but the benefit saturates and then reverses into harm as its share grows, so selective filtering beats blanket inclusion.
• arXiv — ReCAP gives LLM agents a persistent context graph of attention-derived importance scores and dependency links, cutting memory-compaction latency ~95% versus summarization while matching or beating task performance.
• Hugging Face Papers — RIDE distills by extrapolating RL-induced residuals in a model’s hidden-representation space rather than its output space, letting students approach or exceed the RL-trained teacher across every tested pair.
• Hugging Face Papers — EvoDuet co-evolves a web-search agent and a task-solving agent in a bilevel loop, with a “retrieval gate” letting the model decide when to fetch new documents versus reuse what it already has, for scientific-discovery tasks.