News Around the World — 2026-10-20

AI

• Hugging Face Papers — Self-Evolving Coding Agents via Harness-Policy Optimization — Microsoft Research’s Self-HPO alternates letting the agent’s own policy propose harness revisions with RL policy updates under the optimized harness, raising Qwen3.5-35B-A3B’s average coding-benchmark resolve rate from 57.8% to 64.8%.
• arXiv — RefCraft: Benchmarking Visual-Reference-Conditioned Agentic Generation through Web Development — even the strongest of nine frontier multimodal models scores only ~52/100 on building webpages from screenshot references, showing this remains a major weakness for current AI agents.