DeepSeek — Field Evidence
paperUnverified
DeepSeek · DeepSeek · praise
Field note
Large Language Models for Code Generation from Multilingual Prompts: A Curated Benchmark and a Study on Code Quality — Large Language Models (LLMs) perform differently on identical programming tasks when prompted in different natural languages, a phenomenon known as language bias
collected 2026-07-17original 2026-07-16
Does this shift the US–China race?
Be the first to call it
Impact on the index
Benchmark win · minor
China +30US -7
Directional contribution — before recency decay and per-type diminishing returns. How it works →
Related Front
US vs ChinaLikely
U.S. FrontiervsChina Open-Weight
Likely — capability gap narrowing on common tasks
Letters from the Front
Be the first to file a report.