News Around the World — 2026-11-02

AI

• GraphNIAH: A Benchmark for Long-Context Reasoning over Real-World Graphs — On a benchmark built from post-cutoff citation graphs, GPT-5.6 and Qwen3.5 answer YES/NO questions no better than chance at search depths of 16–32, mostly defaulting to NO, suggesting the models stop searching their context before reaching an answer.
• Do Open Language Models Invent Scripture? BILAL-BENCH: Benchmarking Islamic-Source Fabrication in Bengali — Across 240 expert-verified Bengali questions, 6 of 7 open-weight models (1B–120B) confidently fabricate Islamic sources on 79.3–95.3% of source-production questions, and scale doesn’t reliably reduce it.